行业知识
llama2.c
2026-11-05 23:15:52
来源：互联网转载
1、下载模型
从Hugging Face下载��文微型Llama2基础模型，这是一个参数量115M左右的超微型小模型，采用Llama2架构。
（）2、将模型hf格式转换为bin格式
python export.py ./model/chinese-baby-llama2.bin --hf /mnt/workspace/llama2.c/model
model文件夹中命名一个文件chinese-baby-llama2.bin，chinese-baby-llama2压缩包和解压的都放在model文件夹的
3、debug export.py
修改arg中3代码
（）parser.add_argument("--filepath", type=str,default="/mnt/workspace/llama2.c/model/chinese-baby-llama2.bin", help="the output filepath")group = parser.add_mutually_exclusive_group()group.add_argument("--hf", type=str,default="/mnt/workspace/llama2.c/model", help="huggingface model path") 4、几个重点
4.1 model
model = load_hf_model(args.hf)
输出的model
Transformer( (tok_embeddings): Embedding(32000, 768) (dropout): Dropout(p=0.0, inplace=False) (layers): ModuleList( (0): TransformerBlock( (attention): Attention( (wq): Linear(in_features=768, out_features=768, bias=False) (wk): Linear(in_features=768, out_features=768, bias=False) (wv): Linear(in_features=768, out_features=768, bias=False) (wo): Linear(in_features=768, out_features=768, bias=False) (attn_dropout): Dropout(p=0.0, inplace=False) (resid_dropout): Dropout(p=0.0, inplace=False) ) (feed_forward): FeedForward( (w1): Linear(in_features=768, out_features=2268, bias=False) (w2): Linear(in_features=2268, out_features=768, bias=False) (w3): Linear(in_features=768, out_features=2268, bias=False) (dropout): Dropout(p=0.0, inplace=False) ) (attention_norm): RMSNorm() (ffn_norm): RMSNorm() ) (1): TransformerBlock( (attention): Attention( (wq): Linear(in_features=768, out_features=768, bias=False) (wk): Linear(in_features=768, out_features=768, bias=False) (wv): Linear(in_features=768, out_features=768, bias=False) (wo): Linear(in_features=768, out_features=768, bias=False) (attn_dropout): Dropout(p=0.0, inplace=False) (resid_dropout): Dropout(p=0.0, inplace=False) ) (feed_forward): FeedForward( (w1): Linear(in_features=768, out_features=2268, bias=False) (w2): Linear(in_features=2268, out_features=768, bias=False) (w3): Linear(in_features=768, out_features=2268, bias=False) (dropout): Dropout(p=0.0, inplace=False) ) (attention_norm): RMSNorm() (ffn_norm): RMSNorm() ) (2): TransformerBlock( (attention): Attention( (wq): Linear(in_features=768, out_features=768, bias=False) (wk): Linear(in_features=768, out_features=768, bias=False) (wv): Linear(in_features=768, out_features=768, bias=False) (wo): Linear(in_features=768, out_features=768, bias=False) (attn_dropout): Dropout(p=0.0, inplace=False) (resid_dropout): Dropout(p=0.0, inplace=False) ) (feed_forward): FeedForward( (w1): Linear(in_features=768, out_features=2268, bias=False) (w2): Linear(in_features=2268, out_features=768, bias=False) (w3): Linear(in_features=768, out_features=2268, bias=False) (dropout): Dropout(p=0.0, inplace=False) ) (attention_norm): RMSNorm() (ffn_norm): RMSNorm() ) (3): TransformerBlock( (attention): Attention( (wq): Linear(in_features=768, out_features=768, bias=False) (wk): Linear(in_features=768, out_features=768, bias=False) (wv): Linear(in_features=768, out_features=768, bias=False) (wo): Linear(in_features=768, out_features=768, bias=False) (attn_dropout): Dropout(p=0.0, inplace=False) (resid_dropout): Dropout(p=0.0, inplace=False) ) (feed_forward): FeedForward( (w1): Linear(in_features=768, out_features=2268, bias=False) (w2): Linear(in_features=2268, out_features=768, bias=False) (w3): Linear(in_features=768, out_features=2268, bias=False) (dropout): Dropout(p=0.0, inplace=False) ) (attention_norm): RMSNorm() (ffn_norm): RMSNorm() ) (4): TransformerBlock( (attention): Attention( (wq): Linear(in_features=768, out_features=768, bias=False) (wk): Linear(in_features=768, out_features=768, bias=False) (wv): Linear(in_features=768, out_features=768, bias=False) (wo): Linear(in_features=768, out_features=768, bias=False) (attn_dropout): Dropout(p=0.0, inplace=False) (resid_dropout): Dropout(p=0.0, inplace=False) ) (feed_forward): FeedForward( (w1): Linear(in_features=768, out_features=2268, bias=False) (w2): Linear(in_features=2268, out_features=768, bias=False) (w3): Linear(in_features=768, out_features=2268, bias=False) (dropout): Dropout(p=0.0, inplace=False) ) (attention_norm): RMSNorm() (ffn_norm): RMSNorm() ) (5): TransformerBlock( (attention): Attention( (wq): Linear(in_features=768, out_features=768, bias=False) (wk): Linear(in_features=768, out_features=768, bias=False) (wv): Linear(in_features=768, out_features=768, bias=False) (wo): Linear(in_features=768, out_features=768, bias=False) (attn_dropout): Dropout(p=0.0, inplace=False) (resid_dropout): Dropout(p=0.0, inplace=False) ) (feed_forward): FeedForward( (w1): Linear(in_features=768, out_features=2268, bias=False) (w2): Linear(in_features=2268, out_features=768, bias=False) (w3): Linear(in_features=768, out_features=2268, bias=False) (dropout): Dropout(p=0.0, inplace=False) ) (attention_norm): RMSNorm() (ffn_norm): RMSNorm() ) (6): TransformerBlock( (attention): Attention( (wq): Linear(in_features=768, out_features=768, bias=False) (wk): Linear(in_features=768, out_features=768, bias=False) (wv): Linear(in_features=768, out_features=768, bias=False) (wo): Linear(in_features=768, out_features=768, bias=False) (attn_dropout): Dropout(p=0.0, inplace=False) (resid_dropout): Dropout(p=0.0, inplace=False) ) (feed_forward): FeedForward( (w1): Linear(in_features=768, out_features=2268, bias=False) (w2): Linear(in_features=2268, out_features=768, bias=False) (w3): Linear(in_features=768, out_features=2268, bias=False) (dropout): Dropout(p=0.0, inplace=False) ) (attention_norm): RMSNorm() (ffn_norm): RMSNorm() ) (7): TransformerBlock( (attention): Attention( (wq): Linear(in_features=768, out_features=768, bias=False) (wk): Linear(in_features=768, out_features=768, bias=False) (wv): Linear(in_features=768, out_features=768, bias=False) (wo): Linear(in_features=768, out_features=768, bias=False) (attn_dropout): Dropout(p=0.0, inplace=False) (resid_dropout): Dropout(p=0.0, inplace=False) ) (feed_forward): FeedForward( (w1): Linear(in_features=768, out_features=2268, bias=False) (w2): Linear(in_features=2268, out_features=768, bias=False) (w3): Linear(in_features=768, out_features=2268, bias=False) (dropout): Dropout(p=0.0, inplace=False) ) (attention_norm): RMSNorm() (ffn_norm): RMSNorm() ) (8): TransformerBlock( (attention): Attention( (wq): Linear(in_features=768, out_features=768, bias=False) (wk): Linear(in_features=768, out_features=768, bias=False) (wv): Linear(in_features=768, out_features=768, bias=False) (wo): Linear(in_features=768, out_features=768, bias=False) (attn_dropout): Dropout(p=0.0, inplace=False) (resid_dropout): Dropout(p=0.0, inplace=False) ) (feed_forward): FeedForward( (w1): Linear(in_features=768, out_features=2268, bias=False) (w2): Linear(in_features=2268, out_features=768, bias=False) (w3): Linear(in_features=768, out_features=2268, bias=False) (dropout): Dropout(p=0.0, inplace=False) ) (attention_norm): RMSNorm() (ffn_norm): RMSNorm() ) (9): TransformerBlock( (attention): Attention( (wq): Linear(in_features=768, out_features=768, bias=False) (wk): Linear(in_features=768, out_features=768, bias=False) (wv): Linear(in_features=768, out_features=768, bias=False) (wo): Linear(in_features=768, out_features=768, bias=False) (attn_dropout): Dropout(p=0.0, inplace=False) (resid_dropout): Dropout(p=0.0, inplace=False) ) (feed_forward): FeedForward( (w1): Linear(in_features=768, out_features=2268, bias=False) (w2): Linear(in_features=2268, out_features=768, bias=False) (w3): Linear(in_features=768, out_features=2268, bias=False) (dropout): Dropout(p=0.0, inplace=False) ) (attention_norm): RMSNorm() (ffn_norm): RMSNorm() ) (10): TransformerBlock( (attention): Attention( (wq): Linear(in_features=768, out_features=768, bias=False) (wk): Linear(in_features=768, out_features=768, bias=False) (wv): Linear(in_features=768, out_features=768, bias=False) (wo): Linear(in_features=768, out_features=768, bias=False) (attn_dropout): Dropout(p=0.0, inplace=False) (resid_dropout): Dropout(p=0.0, inplace=False) ) (feed_forward): FeedForward( (w1): Linear(in_features=768, out_features=2268, bias=False) (w2): Linear(in_features=2268, out_features=768, bias=False) (w3): Linear(in_features=768, out_features=2268, bias=False) (dropout): Dropout(p=0.0, inplace=False) ) (attention_norm): RMSNorm() (ffn_norm): RMSNorm() ) (11): TransformerBlock( (attention): Attention( (wq): Linear(in_features=768, out_features=768, bias=False) (wk): Linear(in_features=768, out_features=768, bias=False) (wv): Linear(in_features=768, out_features=768, bias=False) (wo): Linear(in_features=768, out_features=768, bias=False) (attn_dropout): Dropout(p=0.0, inplace=False) (resid_dropout): Dropout(p=0.0, inplace=False) ) (feed_forward): FeedForward( (w1): Linear(in_features=768, out_features=2268, bias=False) (w2): Linear(in_features=2268, out_features=768, bias=False) (w3): Linear(in_features=768, out_features=2268, bias=False) (dropout): Dropout(p=0.0, inplace=False) ) (attention_norm): RMSNorm() (ffn_norm): RMSNorm() ) ) (norm): RMSNorm() (output): Linear(in_features=768, out_features=32000, bias=False))
4.2 legacy_export
用于将模型参数以特定格式保存到二进制文件中
4.2.1
out_file.write(header)
构建文件头部信息，其中包括模型的一些参数，如隐藏层维度、层数、注意力头数等。这些参数被打包成一个结构体，并写入到二进制文件中。
4.2.2 serialize_fp32(out_file, model.tok_embeddings.weight)
将模型的 token embeddings 权重写入二进制文件。
4.2.3 循环
for layer in model.layers: serialize_fp32(out_file, layer.attention_norm.weight) for layer in model.layers: serialize_fp32(out_file, layer.attention.wq.weight) for layer in model.layers: serialize_fp32(out_file, layer.attention.wk.weight) for layer in model.layers: serialize_fp32(out_file, layer.attention.wv.weight) for layer in model.layers: serialize_fp32(out_file, layer.attention.wo.weight)
几个循环分别处理模型的注意力层和前馈神经网络层的权重，并将它们写入二进制文件。其中，serialize_fp32用于张量以单精度浮点数的格式写入到二进制文件中。
最终生成chinese-baby-llama2.bin
5、run.c
5.1 文件结构
llama2.c》debug22》CMakeLists.txt
CMakeLists.txt如下：
cmake_minimum_required(VERSION 3.16)project(llama2.c)set(CMAKE_BUILD_TYPE debug) # Debug Releaseset(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_SOURCE_DIR}/")set(CMAKE_CXX_STANDARD 14)SET(CMAKE_C_ FLAGS "${ACMAE_C_FLASS} -O0 -ffast-math -manch=native -fopenmp -mavx2 -mfma -DEISEN_STACK_ALLOCATION_LIMIT=0")SET(CNAKE_CXX_FLAGS "${ACNAKE_CXX_FLASS} -O0 -ffast-math -march=native -fopenmp -mavx2 -mfma -DEITOEN_STACK_ALOCATION_LITRIT=0")add_executable(run /mnt/workspace/llama2.c/run.c)target_link_libraries(run -lpthread -lm -ldl -m64 -lpthread) 5.2 run.c代码
int main(int argc, char *argv[]) { // default parameters char *checkpoint_path = NULL; // e.g. out/model.bin float temperature = 1.0f; // 0.0 = greedy deterministic. 1.0 = original. don't set higher float topp = 0.9f; // top-p in nucleus sampling. 1.0 = off. 0.9 works well, but slower int steps = 256; // number of steps to run for // char *prompt = "NULL"; // prompt string char *prompt = "今天是武林大会，我是武林盟主"; // prompt string unsigned long long rng_seed = 0; // seed rng with time by default char *mode = "generate"; // generate|chat char *system_prompt = NULL; // the (optional) system prompt to use in chat mode // poor man's C argparse so we can override the defaults above from the command line char *tokenizer_path = "/mnt/workspace/llama2.c/model/tokenizer.bin"; if (argc >= 2) { checkpoint_path = argv[1]; } else { error_usage(); } 5.3 开始debug
cd /mnt/workspace/llama2.c/debug22
5.3.1 编译
cmake .make
5.3.2 启动调试器
gdb ./run
5.3.3 设置断点
break main
5.3.3 set args
set args /mnt/workspace/llama2.c/model/chinese-baby-llama2.bin
5.3.4 run起来
run>next == r>n
5.3.5 生成
今天是武林大会，我是武林盟主，也是少林掌门，我们还是拭目以待吧！”一行八人已经呈到击中点，急！这得是多么累的！不适合这华要你为何在这里共操两场，悲催了！刚才给了冷然一剑的地方，就是至于是射中的方向，大厅边缘那处就是形意拳的设计所在，此时却是化作一片虚无。一身黑衣，头上带着一个斗笠的大错睁开眼睛，披肩的长发随意飘舞。左手托着左胸，右手却从小腹下抽出来一只浴桶，不想在这娇美的身躯里被你的小手会碰，让你一下子金星直冒，身躯一阵摇晃，只见又呻吟一声，向后倒去。青石板落下的确是一年多没有有一个骷髅头，最有多年没有进行过了！这一招，还是要用“小之间”，举手投足间，破坏大家的心神！果然，就在千钧一发之际，房顶上也终于全部被铺盖过去了，玩笑，小型的平台，各处都是厚厚的灰，犹如冬日的习惯了，周围更是笼罩着漫天的灰雾，擂台山的边缘处，一片荒凉！如此资质，令欢乐迷醉！随着堂主的暴喝，六百名青衫中年人瞬间起立！九人齐齐的暴体而出，每人
achieved tok/s: 2.399571
参考链接：
①https://zhuanlan.zhihu.com/p/674666408
②https://github.com/karpathy/llama2.c
