昇思25天学习打卡营第16天|LLM-MindNLP ChatGLM-6B StreamChat

UtopXExistential2024-07-22 17:57

打卡

目录

打卡

任务说明

环境配置

部署方式

[ChatGLM-6B 体验截图示例](#ChatGLM-6B 体验截图示例)

[ChatGLM-6B 模型结构解析如下](#ChatGLM-6B 模型结构解析如下)

[ChatGLM2-6B 模型结构解析如下](#ChatGLM2-6B 模型结构解析如下)

任务说明

加载智谱清言的chatglm模型权重文件（目前有4个版本），本次主要尝试了chatglm-6b。

chatglm 6b 在提供的本次实验npu卡中可以正常加载，加载2和3版本时加载tokenizer出错了。但是可以看到model的打印结果，看到chatglm2 和 chatglm3 的模型结构相比1版本，词表扩充了2w+。而我们知道chatglm3-6b 还具有了 functional calling 功能。

环境配置

  • 欧拉操作系统2.0（SP8）；
  • python3.9；
  • 华为 NPU 910A；
  • mindnlp, mindspore, mindvision 安装部署。
  • 其余安装：
python 复制代码 
!pip install mdtex2html
!pip install -i https://pypi.mirrors.ustc.edu.cn/simple mindspore==2.2.14

部署方式

python 复制代码 
from mindnlp.transformers import AutoModelForSeq2SeqLM, AutoTokenizer
import gradio as gr
import mdtex2html


### 下载 && 加载模型权重 
model = AutoModelForSeq2SeqLM.from_pretrained(
            'ZhipuAI/ChatGLM-6B', 
            mirror="modelscope").half()
model.set_train(False)

tokenizer = AutoTokenizer.from_pretrained(
            'ZhipuAI/ChatGLM-6B', 
             mirror="modelscope")

### 修改参数和prompt体验模型
prompt = '你好'
history = []
response, _ = model.chat(tokenizer, prompt, history=history, max_length=20)
response

ChatGLM-6B体验截图示例

如上图，ChatGLM-6B tokenzier 的词表大小是 13w+，有5种特殊的 token。

如下为模型打印结果。

ChatGLM-6B 模型结构解析如下

1. 1级主类： ChatGLMForConditionalGeneration 生成式Transformer模型，用于条件文本生成。

2. 2级类： ChatGLMModel 层，是transformer 结构，是模型的核心部分。

3. 2级类： lm_head 结构的 Dense 全连接层。dim[in, out]=[4096, 130528]

4. ChatGLMModel 结构下的3级类组件分三层：

>> word_embeddings 嵌入层：dim[in, out]=[130528, 4096] ，即使用了 130528 个词汇，每个词汇映射到一个4096维的向量。

>> layers 网络结构层 ：Transformer模型的主体，包含 28 个GLMBlock。

>> final_layernorm最后的层归一化。

5. GLMBlock 的结构：

》》1 input_layernorm层，层归一化，用于在注意力机制之前对输入进行归一化。

》》2 SelfAttention层，自注意力机制，用于计算输入序列中不同位置的注意力权重。共包括3层：RotaryEmbedding 旋转嵌入，用于增强模型对位置信息的处理能力； Dense（query_key_value）用于生成查询（Q）、键（K）和值（V）向量；Dense（Dense）用于自注意力机制的输出。。

》》3 post_attention_layernorm层，用于自注意力之后的归一化。

》》4 mlp 层 ，多层感知机，用于对自注意力层的输出进行进一步的非线性变换。这里的MLP使用的是GLU（门控线性单元）激活函数。dense_h_to_4h 线性变换将输入维度扩大4倍。dense_4h_to_h 将扩大后的维度还原。

如下图，chatglm-6b model 的打印结果。

python 复制代码 
$ print(model)

ChatGLMForConditionalGeneration<
  (transformer): ChatGLMModel<
    (word_embeddings): Embedding<vocab_size=130528, embedding_size=4096, use_one_hot=False, weight=Parameter (Tensor(shape=[130528, 4096], dtype=Float16, value=[...], name=transformer.word_embeddings.weight), requires_grad=True), dtype=Float16, padding_idx=None>
    (layers): CellList<
      (0): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.0.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.0.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.0.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.0.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (1): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.1.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.1.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.1.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.1.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (2): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.2.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.2.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.2.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.2.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (3): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.3.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.3.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.3.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.3.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (4): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.4.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.4.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.4.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.4.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (5): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.5.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.5.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.5.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.5.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (6): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.6.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.6.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.6.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.6.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (7): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.7.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.7.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.7.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.7.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (8): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.8.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.8.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.8.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.8.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (9): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.9.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.9.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.9.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.9.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (10): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.10.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.10.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.10.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.10.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (11): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.11.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.11.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.11.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.11.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (12): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.12.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.12.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.12.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.12.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (13): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.13.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.13.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.13.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.13.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (14): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.14.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.14.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.14.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.14.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (15): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.15.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.15.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.15.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.15.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (16): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.16.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.16.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.16.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.16.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (17): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.17.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.17.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.17.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.17.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (18): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.18.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.18.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.18.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.18.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (19): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.19.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.19.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.19.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.19.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (20): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.20.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.20.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.20.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.20.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (21): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.21.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.21.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.21.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.21.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (22): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.22.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.22.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.22.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.22.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (23): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.23.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.23.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.23.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.23.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (24): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.24.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.24.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.24.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.24.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (25): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.25.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.25.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.25.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.25.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (26): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.26.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.26.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.26.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.26.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (27): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.27.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.27.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.27.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.27.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      >
    (final_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.final_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.final_layernorm.bias), requires_grad=True)>
    >
  (lm_head): Dense<input_channels=4096, output_channels=130528>
  >

ChatGLM2-6B 模型结构解析如下

如下图，chatglm2-6b model 的打印结果。相比于1版本，模型结构没有变化，只是vocab_size词表扩充成了15w+。

python 复制代码 
$ print(model)


ChatGLMForConditionalGeneration<
  (transformer): ChatGLMModel<
    (word_embeddings): Embedding<vocab_size=150528, embedding_size=4096, use_one_hot=False, weight=Parameter (Tensor(shape=[150528, 4096], dtype=Float16, value=[...], name=transformer.word_embeddings.weight), requires_grad=True), dtype=Float16, padding_idx=None>
    (layers): CellList<
      (0): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.0.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.0.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.0.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.0.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (1): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.1.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.1.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.1.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.1.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (2): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.2.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.2.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.2.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.2.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (3): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.3.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.3.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.3.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.3.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (4): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.4.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.4.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.4.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.4.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (5): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.5.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.5.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.5.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.5.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (6): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.6.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.6.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.6.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.6.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (7): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.7.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.7.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.7.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.7.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (8): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.8.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.8.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.8.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.8.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (9): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.9.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.9.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.9.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.9.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (10): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.10.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.10.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.10.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.10.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (11): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.11.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.11.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.11.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.11.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (12): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.12.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.12.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.12.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.12.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (13): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.13.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.13.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.13.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.13.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (14): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.14.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.14.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.14.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.14.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (15): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.15.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.15.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.15.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.15.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (16): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.16.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.16.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.16.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.16.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (17): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.17.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.17.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.17.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.17.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (18): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.18.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.18.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.18.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.18.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (19): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.19.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.19.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.19.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.19.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (20): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.20.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.20.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.20.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.20.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (21): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.21.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.21.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.21.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.21.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (22): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.22.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.22.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.22.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.22.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (23): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.23.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.23.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.23.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.23.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (24): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.24.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.24.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.24.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.24.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (25): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.25.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.25.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.25.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.25.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (26): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.26.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.26.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.26.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.26.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      (27): GLMBlock<
        (input_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.27.input_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.27.input_layernorm.bias), requires_grad=True)>
        (attention): SelfAttention<
          (rotary_emb): RotaryEmbedding<>
          (query_key_value): Dense<input_channels=4096, output_channels=12288, has_bias=True>
          (dense): Dense<input_channels=4096, output_channels=4096, has_bias=True>
          >
        (post_attention_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.27.post_attention_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.layers.27.post_attention_layernorm.bias), requires_grad=True)>
        (mlp): GLU<
          (dense_h_to_4h): Dense<input_channels=4096, output_channels=16384, has_bias=True>
          (dense_4h_to_h): Dense<input_channels=16384, output_channels=4096, has_bias=True>
          >
        >
      >
    (final_layernorm): LayerNorm<normalized_shape=[4096], begin_norm_axis=-1, begin_params_axis=-1, weight=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.final_layernorm.weight), requires_grad=True), bias=Parameter (Tensor(shape=[4096], dtype=Float16, value=[...], name=transformer.final_layernorm.bias), requires_grad=True)>
    >
  (lm_head): Dense<input_channels=4096, output_channels=150528>
  >
相关推荐
晓数1 小时前
“平价”微智码初尝试
人工智能·jetbrains
新加坡内哥谈技术1 小时前
MCP：人工智能时代的HTTP？探索AI通信新标准
人工智能·自然语言处理·chatgpt
0x2113 小时前
[论文阅读]REPLUG: Retrieval-Augmented Black-Box Language Models
论文阅读·人工智能·语言模型
JOYCE_Leo164 小时前
一文详解卷积神经网络中的卷积层和池化层原理 ！！
人工智能·深度学习·cnn·卷积神经网络
～央千澈～6 小时前
对鸿蒙 Next 系统“成熟论”的深度剖析-优雅草卓伊凡
人工智能
Donvink6 小时前
【视频生成模型】通义万相Wan2.1模型本地部署和LoRA微调
人工智能·深度学习·aigc·音视频
訾博ZiBo6 小时前
AI日报 - 2025年04月29日
人工智能
爱喝奶茶的企鹅6 小时前
Ethan独立开发产品日报 | 2025-04-27
人工智能·程序员·开源
极小狐6 小时前
如何对极狐GitLab 议题进行过滤和排序？
人工智能·git·机器学习·gitlab
热门推荐
01西电B测-计算机网络综合实验(含验收问题)02KGG转MP3工具|非KGM文件|解密音频03从零安装 LLaMA-Factory 微调 Qwen 大模型成功及所有的坑04Coze扣子平台完整体验和实践（附国内和国际版对比）05YOLOv8入门 | 重要性能衡量指标、训练结果评价及分析及影响mAP的因素【发论文关注的指标】06DeepSeek各版本说明与优缺点分析07yolov10来了！用yolov10训练自己的数据集（原理、训练、部署、应用）08yolov8，yolo11，yolo12 服务器训练到部署全流程 笔记09我决定放弃搞 Java 了10最新 Kubernetes 集群部署 + flannel 网络插件（保姆级教程，最新 K8S 版本）