{"id":3869,"date":"2024-05-21T19:06:09","date_gmt":"2024-05-21T11:06:09","guid":{"rendered":"https:\/\/www.aqwu.net\/wp\/?p=3869"},"modified":"2024-05-22T21:47:48","modified_gmt":"2024-05-22T13:47:48","slug":"llama3-%e4%bb%8e%e5%a4%b4%e5%bc%80%e5%a7%8b%e5%ae%9e%e7%8e%b0","status":"publish","type":"post","link":"https:\/\/www.aqwu.net\/wp\/?p=3869","title":{"rendered":"llama3 \u4ece\u5934\u5f00\u59cb\u5b9e\u73b0"},"content":{"rendered":"\n

\u539f\u6587\u53c2\u8003\u8fde\u63a5\uff1ahttps:\/\/github.com\/naklecha\/llama3-from-scratch<\/a><\/p>\n\n\n\n

\u5728\u672c\u7bc7\u6587\u7ae0\u4e2d\uff0c\u4ece\u5934\u5f00\u59cb\u5b9e\u73b0\u4e86 llama3\uff0c\u4e00\u6b21\u4e00\u4e2a\u5f20\u91cf\u548c\u77e9\u9635\u4e58\u6cd5\u3002<\/p>\n\n\n\n

\u6b64\u5916\uff0c\u5c06\u76f4\u63a5\u4ece Meta \u4e3a llama3 \u63d0\u4f9b\u7684\u6a21\u578b\u6587\u4ef6\u52a0\u8f7d\u5f20\u91cf\uff0c\u60a8\u9700\u8981\u5728\u8fd0\u884c\u6b64\u6587\u4ef6\u4e4b\u524d\u4e0b\u8f7d\u6743\u91cd\u3002\u4ee5\u4e0b\u662f\u4e0b\u8f7dllama3\u6743\u91cd\u7684\u5b98\u65b9\u94fe\u63a5\uff1a https:\/\/llama.meta.com\/llama-downloads\/<\/p>\n\n\n\n

\u5b9e\u9a8c\u662f\u5728 Windows 11 \u73af\u5883\u8fdb\u884c\u7684\uff0c\u4f7f\u7528 jupyter notebook \u8fdb\u884c\u8c03\u8bd5<\/p>\n\n\n\n

1. \u521b\u5efa jupyter \u73af\u5883<\/strong><\/h2>\n\n\n\n
conda create -n jupyter_env python==3.11\nconda activate jupyter_env\n\nconda install -c conda-forge notebook\nconda install -c conda-forge jupyter_core\n<\/pre><\/div>\n\n\n\n
1.1 \u786e\u8ba4 python.exe \u8def\u5f84<\/strong><\/h3>\n\n\n\n
where python.exe\nC:\\Users\\tony1\\anaconda3\\envs\\jupyter_env\\python.exe<\/pre><\/div>\n\n\n\n
1.2 \u786e\u8ba4 pip.exe \u8def\u5f84<\/strong><\/h3>\n\n\n\n
where pip.exe\nC:\\Users\\tony1\\anaconda3\\envs\\jupyter_env\\Scripts\\pip.exe<\/pre><\/div>\n\n\n\n
1.3 \u786e\u8ba4 jupyter.exe \u8def\u5f84<\/strong><\/h3>\n\n\n\n
where jupyter.exe\nC:\\Users\\tony1\\anaconda3\\envs\\jupyter_env\\Scripts\\jupyter.exe<\/pre><\/div>\n\n\n\n
1.4 \u8fd0\u884c notebook<\/strong><\/h3>\n\n\n\n
jupyter notebook\n<\/pre><\/div>\n\n\n\n
2. \u5b89\u88c5\u5fc5\u8981\u7684\u5e93<\/strong><\/h2>\n\n\n\n
!pip install sentencepiece\n!pip install tiktoken\n!pip install torch\n!pip install blobfile\n!pip install matplotlib<\/pre><\/div>\n\n\n\n
3. \u5206\u8bcd\u5668(tokenizer)<\/strong><\/h2>\n\n\n\n
\u8fd9\u91cc\u4e0d\u6253\u7b97\u5b9e\u73b0 BPE \u5206\u8bcd\u5668\uff08\u4f46 Andrej Karpathy \u6709\u4e00\u4e2a\u975e\u5e38\u5e72\u51c0\u7684\u5b9e\u73b0\uff09<\/p>\n\n\n\n
https:\/\/github.com\/karpathy\/minbpe<\/a><\/p>\n\n\n\n
\u53ef\u4ee5\u4ece llama3 \u7684 tokenizer.json \u6587\u4ef6\u77e5\u9053\u5206\u8bcd\u5668\u7c7b\u578b\u662f BPE<\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\u5728\u6587\u4ef6 tokenizer.json \u7684\u524d\u9762\u8fd8\u6709\u5176\u4ed6\u7279\u6b8a\u7684 tokens<\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\u4e0b\u9762\u7684\u4ee3\u7801\u542b\u6709\u4e86 tokenizer.json \u6587\u4ef6\u91cc\u9762\u6240\u6709\u7279\u5b9a\u7684\u5b57\u7b26<\/p>\n\n\n\n
from pathlib import Path\nimport tiktoken\nfrom tiktoken.load import load_tiktoken_bpe\nimport torch\nimport json\nimport matplotlib.pyplot as plt\n\ntokenizer_path = \"Meta-Llama-3-8B\/tokenizer.model\"\nspecial_tokens = [\n            \"<|begin_of_text|>\",\n            \"<|end_of_text|>\",\n            \"<|reserved_special_token_0|>\",\n            \"<|reserved_special_token_1|>\",\n            \"<|reserved_special_token_2|>\",\n            \"<|reserved_special_token_3|>\",\n            \"<|start_header_id|>\",\n            \"<|end_header_id|>\",\n            \"<|reserved_special_token_4|>\",\n            \"<|eot_id|>\",  # end of turn\n        ] + [f\"<|reserved_special_token_{i}|>\" for i in range(5, 256 - 5)]\nmergeable_ranks = load_tiktoken_bpe(tokenizer_path)\ntokenizer = tiktoken.Encoding(\n    name=Path(tokenizer_path).name,\n    pat_str=r\"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+\",\n    mergeable_ranks=mergeable_ranks,\n    special_tokens={token: len(mergeable_ranks) + i for i, token in enumerate(special_tokens)},\n)\n\nprint(tokenizer.n_vocab)\nprint(tokenizer.decode(tokenizer.encode(\"hello world!\")))<\/pre><\/div>\n\n\n\n
\u8fd0\u884c\u7ed3\u679c\uff1a<\/p>\n\n\n\n
128256\nhello world!<\/pre><\/div>\n\n\n\n
\u53ef\u4ee5\u770b\u5230 tokenizer.n_vocab \u7684\u503c\u4e3a 128256\uff0c\u548c config.json \u6587\u4ef6\u91cc\u7684 vocab_size \u4e00\u81f4<\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
4. \u8bfb\u53d6\u6a21\u578b\u6587\u4ef6<\/strong><\/h2>\n\n\n\n
\u4ece\u5934\u5f00\u59cb\u5b9e\u73b0 llama3\uff0c\u6211\u4eec\u5c06\u4e00\u6b21\u8bfb\u53d6\u4e00\u4e2a\u5f20\u91cf\u6587\u4ef6\u3002<\/p>\n\n\n\n
model = torch.load(\"Meta-Llama-3-8B\/consolidated.00.pth\")\n# Print the first 20 keys from the model dictionary\nprint(json.dumps(list(model.keys())[:20], indent=4))\n\n# Print the last two keys from the model dictionary\nprint(json.dumps(list(model.keys())[-2:], indent=4))<\/pre><\/div>\n\n\n\n
\u8fd0\u884c\u7ed3\u679c\uff1a<\/p>\n\n\n\n
[\n    \"tok_embeddings.weight\",\n    \"layers.0.attention.wq.weight\",\n    \"layers.0.attention.wk.weight\",\n    \"layers.0.attention.wv.weight\",\n    \"layers.0.attention.wo.weight\",\n    \"layers.0.feed_forward.w1.weight\",\n    \"layers.0.feed_forward.w3.weight\",\n    \"layers.0.feed_forward.w2.weight\",\n    \"layers.0.attention_norm.weight\",\n    \"layers.0.ffn_norm.weight\",\n    \"layers.1.attention.wq.weight\",\n    \"layers.1.attention.wk.weight\",\n    \"layers.1.attention.wv.weight\",\n    \"layers.1.attention.wo.weight\",\n    \"layers.1.feed_forward.w1.weight\",\n    \"layers.1.feed_forward.w3.weight\",\n    \"layers.1.feed_forward.w2.weight\",\n    \"layers.1.attention_norm.weight\",\n    \"layers.1.ffn_norm.weight\",\n    \"layers.2.attention.wq.weight\"\n]\n[\n    \"norm.weight\",\n    \"output.weight\"\n]<\/pre><\/div>\n\n\n\n
\u8fd9\u4e9b\u952e\u540d\u4ee3\u8868\u4e86\u6a21\u578b\u4e2d\u4e0d\u540c\u90e8\u4ef6\u7684\u53c2\u6570\u3002\u4ee5\u4e0b\u662f\u5bf9\u6bcf\u4e2a\u952e\u7684\u89e3\u91ca\uff1a<\/p>\n\n\n\n
    \n
  1. tok_embeddings.weight<\/strong> – \u4ee3\u8868\u8bcd\u5d4c\u5165\u5c42\u7684\u6743\u91cd\uff0c\u5b83\u8d1f\u8d23\u5c06\u8f93\u5165\u7684\u5355\u8bcd\uff08\u6216\u6807\u8bb0\uff09\u8f6c\u6362\u4e3a\u56fa\u5b9a\u5927\u5c0f\u7684\u5411\u91cf\u3002<\/li>\n\n\n\n
  2. layers.x.attention.wq.weight, wk.weight, wv.weight, wo.weight<\/strong> – \u8fd9\u4e9b\u5206\u522b\u8868\u793a\u5728Transformer\u7684\u7b2cx\u5c42\u4e2d\uff0c\u591a\u5934\u6ce8\u610f\u529b\uff08multi-head attention\uff09\u673a\u5236\u7684\u67e5\u8be2\uff08Query\uff09\u3001\u952e\uff08Key\uff09\u3001\u503c\uff08Value\uff09\u548c\u8f93\u51fa\uff08Output\uff09\u7684\u6743\u91cd\u3002<\/li>\n\n\n\n
  3. layers.x.feed_forward.w1.weight, w2.weight, w3.weight<\/strong> – \u8fd9\u4e9b\u4ee3\u8868\u5728\u7b2cx\u5c42\u7684\u524d\u9988\u795e\u7ecf\u7f51\u7edc\uff08feed-forward network, FFN\uff09\u4e2d\u7684\u4e0d\u540c\u6743\u91cd\u3002\u901a\u5e38\u5728BERT\u6216Transformer\u6a21\u578b\u4e2d\uff0cFFN\u5305\u542b\u4e24\u4e2a\u7ebf\u6027\u53d8\u6362\uff0c\u8fd9\u91ccw1\u548cw2\u53ef\u80fd\u4ee3\u8868\u8fd9\u4e24\u4e2a\u53d8\u6362\u7684\u6743\u91cd\uff0cw3\u53ef\u80fd\u662f\u6269\u5c55\u5b9e\u73b0\u7684\u4e00\u90e8\u5206\u3002<\/li>\n\n\n\n
  4. layers.x.attention_norm.weight, layers.x.ffn_norm.weight<\/strong> – \u8fd9\u4e9b\u662f\u5e94\u7528\u5728\u591a\u5934\u6ce8\u610f\u529b\u548c\u524d\u9988\u795e\u7ecf\u7f51\u7edc\u8f93\u51fa\u4e0a\u7684\u5c42\u5f52\u4e00\u5316\uff08layer normalization\uff09\u7684\u6743\u91cd\u3002<\/li>\n\n\n\n
  5. norm.weight<\/strong> – \u8fd9\u53ef\u80fd\u662f\u6a21\u578b\u6700\u540e\u7684\u5c42\u5f52\u4e00\u5316\u6743\u91cd\uff0c\u7528\u4e8e\u6574\u4e2a\u6a21\u578b\u8f93\u51fa\u7684\u6700\u7ec8\u5f52\u4e00\u5316\u5904\u7406\u3002<\/li>\n\n\n\n
  6. output.weight<\/strong> – \u8fd9\u662f\u6a21\u578b\u8f93\u51fa\u5c42\u7684\u6743\u91cd\u3002<\/li>\n<\/ol>\n\n\n\n
    \u4ee5\u4e0a\u6bcf\u4e2a\u53c2\u6570\u90fd\u5728\u6a21\u578b\u5185\u90e8\u627f\u62c5\u7740\u91cd\u8981\u7684\u89d2\u8272\uff0c\u786e\u4fdd\u4fe1\u606f\u80fd\u591f\u5728\u6a21\u578b\u5c42\u95f4\u6709\u6548\u4f20\u9012\u5e76\u8fdb\u884c\u9002\u5f53\u7684\u53d8\u6362\u3002<\/p>\n\n\n\n
    \u4e0b\u9762\u662f\u67e5\u770b\u5176\u4ed6\u6a21\u578b\u53c2\u6570\uff1a<\/p>\n\n\n\n
    with open(\"Meta-Llama-3-8B\/params.json\", \"r\") as f:\n    config = json.load(f)\nconfig<\/pre><\/div>\n\n\n\n
    \u8fd0\u884c\u7ed3\u679c\uff1a<\/p>\n\n\n\n
    {'dim': 4096,\n 'n_layers': 32,\n 'n_heads': 32,\n 'n_kv_heads': 8,\n 'vocab_size': 128256,\n 'multiple_of': 1024,\n 'ffn_dim_multiplier': 1.3,\n 'norm_eps': 1e-05,\n 'rope_theta': 500000.0}<\/pre><\/div>\n\n\n\n
    \u8fd9\u4e9b\u53c2\u6570\u662f\u7528\u4e8e\u914d\u7f6e\u4e00\u4e2a\u57fa\u4e8eTransformer\u67b6\u6784\u7684\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u7684\u5404\u79cd\u8bbe\u7f6e\u3002\u6bcf\u4e2a\u53c2\u6570\u7684\u5177\u4f53\u4f5c\u7528\u5982\u4e0b\uff1a<\/p>\n\n\n\n
      \n
    1. dim<\/strong>: \u8868\u793a\u6a21\u578b\u4e2d\u6bcf\u5c42\u7684\u7ef4\u5ea6\u6216\u9690\u85cf\u5c42\u7684\u5927\u5c0f\u3002\u5728\u8fd9\u91cc\uff0c\u5b83\u88ab\u8bbe\u7f6e\u4e3a4096\uff0c\u610f\u5473\u7740\u6bcf\u4e2aTransformer\u5c42\u7684\u5185\u90e8\u5411\u91cf\u5c06\u67094096\u4e2a\u7ef4\u5ea6\u3002<\/li>\n\n\n\n
    2. n_layers<\/strong>: \u6a21\u578b\u4e2d\u7684\u5c42\u6570\u3002\u8fd9\u4e2a\u53c2\u6570\u503c\u4e3a32\uff0c\u610f\u5473\u7740\u8fd9\u4e2a\u6a21\u578b\u670932\u5c42Transformer\u5c42\u3002<\/li>\n\n\n\n
    3. n_heads<\/strong>: \u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\u4e2d\u5934\u7684\u6570\u91cf\u3002\u5728\u8fd9\u91cc\uff0c\u5b83\u662f32\uff0c\u8868\u793a\u6bcf\u4e2a\u6ce8\u610f\u529b\u5c42\u4f1a\u670932\u4e2a\u4e0d\u540c\u7684\u5934\u5e76\u884c\u5904\u7406\u4fe1\u606f\u3002<\/li>\n\n\n\n
    4. n_kv_heads<\/strong>: \u7279\u5b9a\u4e8e\u67d0\u4e9b\u6a21\u578b\u7684\u53c2\u6570\uff0c\u53ef\u80fd\u8868\u793a\u5728\u5904\u7406\u952e\uff08K\uff09\u548c\u503c\uff08V\uff09\u65f6\u4f7f\u7528\u7684\u6ce8\u610f\u529b\u5934\u7684\u6570\u91cf\u3002\u5728\u8fd9\u4e2a\u6a21\u578b\u4e2d\uff0c\u8fd9\u4e2a\u6570\u503c\u4e3a8\u3002<\/li>\n\n\n\n
    5. vocab_size<\/strong>: \u8bcd\u6c47\u8868\u7684\u5927\u5c0f\uff0c\u5373\u6a21\u578b\u53ef\u4ee5\u8bc6\u522b\u7684\u4e0d\u540c\u5355\u8bcd\u6216\u6807\u8bb0\u7684\u603b\u6570\u3002\u8fd9\u91cc\u8bbe\u7f6e\u4e3a128,256\u3002<\/li>\n\n\n\n
    6. multiple_of<\/strong>: \u8fd9\u901a\u5e38\u7528\u4e8e\u786e\u4fdd\u6a21\u578b\u7ef4\u5ea6\u7684\u4e00\u4e9b\u65b9\u9762\uff08\u5982\u5c42\u5927\u5c0f\uff09\u662f\u67d0\u4e2a\u6570\u503c\uff081024\uff09\u7684\u500d\u6570\uff0c\u8fd9\u6709\u52a9\u4e8e\u4f18\u5316\u8ba1\u7b97\u6548\u7387\u3002<\/li>\n\n\n\n
    7. ffn_dim_multiplier<\/strong>: \u524d\u9988\u7f51\u7edc\u7ef4\u5ea6\u4e58\u6570\u3002\u8fd9\u91cc\u662f1.3\uff0c\u610f\u5473\u7740\u524d\u9988\u5c42\u7684\u5185\u90e8\u7ef4\u5ea6\u662f\u8f93\u5165\/\u8f93\u51fa\u7ef4\u5ea6\uff08dim\uff09\u76841.3\u500d\uff0c\u5373\u8ba1\u7b97\u4e3a (4096 \\times 1.3).<\/li>\n\n\n\n
    8. norm_eps<\/strong>: \u5c42\u5f52\u4e00\u5316\u64cd\u4f5c\u4e2d\u4f7f\u7528\u7684\u4e00\u4e2a\u6781\u5c0f\u6570\uff08epsilon\uff09\uff0c\u7528\u4e8e\u907f\u514d\u9664\u4ee5\u96f6\u7684\u9519\u8bef\u3002\u8fd9\u91cc\u8bbe\u7f6e\u4e3a (10^{-5})\u3002<\/li>\n\n\n\n
    9. rope_theta<\/strong>: \u7279\u5b9a\u4e8e\u67d0\u4e9bTransformer\u6a21\u578b\u7684\u53c2\u6570\uff0c\u53ef\u80fd\u4e0e\u65cb\u8f6c\u4f4d\u7f6e\u7f16\u7801\uff08RoPE\uff09\u6709\u5173\u3002\u5728\u8fd9\u91cc\uff0c\u5b83\u8bbe\u7f6e\u4e3a500000.0\uff0c\u8fd9\u53ef\u80fd\u5f71\u54cd\u6a21\u578b\u5904\u7406\u4f4d\u7f6e\u4fe1\u606f\u7684\u65b9\u5f0f\u3002<\/li>\n<\/ol>\n\n\n\n
      \u8fd9\u4e9b\u53c2\u6570\u5171\u540c\u5b9a\u4e49\u4e86\u4e00\u4e2a\u590d\u6742\u7684\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u7684\u7ed3\u6784\u548c\u884c\u4e3a\uff0c\u6bcf\u4e2a\u90fd\u5bf9\u6a21\u578b\u7684\u6027\u80fd\u548c\u80fd\u529b\u4ea7\u751f\u91cd\u8981\u5f71\u54cd\u3002<\/p>\n\n\n\n
      \u4fdd\u5b58\u8fd9\u4e9b\u53c2\u6570<\/strong><\/p>\n\n\n\n
      dim = config[\"dim\"]\nn_layers = config[\"n_layers\"]\nn_heads = config[\"n_heads\"]\nn_kv_heads = config[\"n_kv_heads\"]\nvocab_size = config[\"vocab_size\"]\nmultiple_of = config[\"multiple_of\"]\nffn_dim_multiplier = config[\"ffn_dim_multiplier\"]\nnorm_eps = config[\"norm_eps\"]\nrope_theta = torch.tensor(config[\"rope_theta\"])<\/pre><\/div>\n\n\n\n
      \u8fd9\u4e9b\u884c\u4ececonfig<\/code>\u7684\u5b57\u5178\u4e2d\u63d0\u53d6\u7279\u5b9a\u7684\u6a21\u578b\u914d\u7f6e\u53c2\u6570\u3002\u8fd9\u5305\u62ec\u6a21\u578b\u7684\u7ef4\u5ea6\u3001\u5c42\u6570\u3001\u5934\u7684\u6570\u91cf\u7b49\uff0c\u4ee5\u53ca\u5c06rope_theta<\/code>\u53c2\u6570\u8f6c\u6362\u4e3a\u4e00\u4e2aPyTorch\u5f20\u91cf\u3002<\/p>\n\n\n\n
      5. \u5c06\u6587\u672c\u8f6c\u6362\u4e3a\u6807\u8bb0<\/strong><\/h2>\n\n\n\n
      \u5728\u8fd9\u91cc\uff0c\u6211\u4eec\u4f7f\u7528 tiktoken\uff08\u6211\u8ba4\u4e3a\u662f OpenAI \u5e93\uff09\u4f5c\u4e3a\u5206\u8bcd\u5668<\/p>\n\n\n\n
      prompt = \"the answer to the ultimate question of life, the universe, and everything is \"\ntokens = [128000] + tokenizer.encode(prompt)\nprint(tokens)\nprint(len(tokens))<\/pre><\/div>\n\n\n\n
      \u8fd0\u884c\u7ed3\u679c\uff1a<\/p>\n\n\n\n
      [128000, 1820, 4320, 311, 279, 17139, 3488, 315, 2324, 11, 279, 15861, 11, 323, 4395, 374, 220]\n17<\/pre><\/div>\n\n\n\n
      \u8fd9\u91cc 128000<\/strong>\uff0c\u4ee3\u8868 bos_token_id\uff0cbos_token\uff0c”<|begin_of_text|>”<\/p>\n\n\n\n
      17<\/strong> \u8868\u793atokens \u7684\u957f\u5ea6\u3002<\/p>\n\n\n\n
      \u8fd9\u91cc\u5b9a\u4e49\u4e86\u4e00\u4e2a\u6587\u672c\u63d0\u793a\uff0c\u5e76\u4f7f\u7528tokenizer.encode<\/code>\u65b9\u6cd5\u5c06\u5176\u8f6c\u6362\u4e3a\u6807\u8bb0\u3002\u6807\u8bb0128000<\/code>\u88ab\u6dfb\u52a0\u5230\u7f16\u7801\u7684\u6807\u8bb0\u5217\u8868\u7684\u5f00\u5934\uff0c\u53ef\u80fd\u4ee3\u8868\u67d0\u79cd\u7279\u6b8a\u6216\u8d77\u59cb\u6807\u8bb0\u3002\u63a5\u7740\u6253\u5370\u8fd9\u4e9b\u6807\u8bb0\u53ca\u5176\u6570\u91cf\u3002<\/p>\n\n\n\n
      \u6211\u4eec\u518d\u4ecetokens \u8fd8\u539f\u51fa\u6765\u539f\u6765\u7684\u5185\u5bb9:<\/p>\n\n\n\n
      tokens = torch.tensor(tokens)\nprompt_split_as_tokens = [tokenizer.decode([token.item()]) for token in tokens]\nprint(prompt_split_as_tokens)<\/pre><\/div>\n\n\n\n
      \u8fd0\u884c\u7ed3\u679c\uff1a<\/p>\n\n\n\n
      ['<|begin_of_text|>', 'the', ' answer', ' to', ' the', ' ultimate', ' question', ' of', ' life', ',', ' the', ' universe', ',', ' and', ' everything', ' is', ' ']<\/pre><\/div>\n\n\n\n
      \u5c06\u6807\u8bb0\u5217\u8868\u8f6c\u6362\u4e3aPyTorch\u5f20\u91cf\uff0c\u7136\u540e\u8fed\u4ee3\u6bcf\u4e2a\u6807\u8bb0\uff0c\u5c06\u5176\u89e3\u7801\u56de\u6587\u672c\u5f62\u5f0f\u5e76\u6253\u5370\u3002\u8fd9\u662f\u4e3a\u4e86\u9a8c\u8bc1\u6807\u8bb0\u662f\u5426\u6b63\u786e\u89e3\u7801\u56de\u5176\u539f\u59cb\u6587\u672c\u3002<\/p>\n\n\n\n
      6. \u5c06\u4ee4\u724c\u8f6c\u6362\u4e3a\u5176\u5d4c\u5165<\/strong><\/h2>\n\n\n\n
      \u5bf9\u4e0d\u8d77\uff0c\u8fd9\u662f\u4f7f\u7528\u5185\u7f6e\u795e\u7ecf\u7f51\u7edc\u6a21\u5757\u7684\u4ee3\u7801\u5e93\u4e2d\u552f\u4e00\u90e8\u5206<\/p>\n\n\n\n
      \u65e0\u8bba\u5982\u4f55\uff0c\u6211\u4eec\u7684 [17×1] \u4ee4\u724c\u73b0\u5728\u662f [17×4096]\uff0c\u5373 17 \u4e2a\u957f\u5ea6\u4e3a 4096 \u7684\u5d4c\u5165\uff08\u6bcf\u4e2a\u4ee4\u724c\u4e00\u4e2a\uff09<\/p>\n\n\n\n
      embedding_layer = torch.nn.Embedding(vocab_size, dim)\nembedding_layer.weight.data.copy_(model[\"tok_embeddings.weight\"])\ntoken_embeddings_unnormalized = embedding_layer(tokens).to(torch.bfloat16)\ntoken_embeddings_unnormalized.shape\nprint(token_embeddings_unnormalized)\n<\/pre><\/div>\n\n\n\n
      \u8fd0\u884c\u7ed3\u679c\uff1a<\/p>\n\n\n\n
      torch.Size([17, 4096])<\/pre><\/div>\n\n\n\n
      tensor([[-1.4305e-04,  1.0777e-04, -1.9646e-04,  ...,  2.0218e-04,\n          1.4842e-05,  3.0136e-04],\n        [ 1.4267e-03, -5.1575e-03,  2.0905e-03,  ...,  1.0437e-02,\n         -9.7656e-03, -7.8735e-03],\n        [-1.1719e-02,  4.4861e-03,  1.1444e-03,  ...,  9.0790e-04,\n          1.0681e-02, -6.7444e-03],\n        ...,\n        [-4.3945e-03, -6.2866e-03,  9.8877e-03,  ...,  5.1575e-03,\n          2.3193e-03, -8.6212e-04],\n        [-2.9144e-03,  1.5335e-03, -2.4605e-04,  ...,  9.7046e-03,\n          4.2114e-03,  7.9956e-03],\n        [-2.1172e-04, -1.5545e-04,  1.5945e-03,  ..., -1.6632e-03,\n          1.1673e-03, -1.1826e-03]], dtype=torch.bfloat16,\n       grad_fn=<ToCopyBackward0>)<\/pre><\/div>\n\n\n\n
      \u8fd9\u90e8\u5206\u4ee3\u7801\u521b\u5efa\u4e86\u4e00\u4e2a\u8bcd\u5d4c\u5165\u5c42\uff0c\u5176\u5927\u5c0f\u4e3avocab_size<\/code> x dim<\/code>\u3002\u7136\u540e\u5c06\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u8bcd\u5d4c\u5165\u6743\u91cd\u590d\u5236\u5230\u8fd9\u4e2a\u65b0\u5c42\u7684\u6743\u91cd\u4e2d\u3002\u4f7f\u7528\u8fd9\u4e2a\u8bcd\u5d4c\u5165\u5c42\u5c06\u8f93\u5165\u7684\u6807\u8bb0\u8f6c\u6362\u4e3a\u5d4c\u5165\u5411\u91cf\uff0c\u5e76\u5c06\u6570\u636e\u7c7b\u578b\u8f6c\u6362\u4e3abfloat16<\/code>\uff08\u4e00\u79cd\u7528\u4e8e\u6df1\u5ea6\u5b66\u4e60\u4e2d\u8282\u7701\u5185\u5b58\u7684\u6d6e\u70b9\u683c\u5f0f\uff09\u3002\u6700\u540e\uff0c\u6253\u5370\u5d4c\u5165\u5411\u91cf\u7684\u5f62\u72b6(shapes)\u548c\u5185\u5bb9\u3002<\/p>\n\n\n\n
      7. \u4f7f\u7528 RMS \u89c4\u8303\u5316\u5bf9\u5d4c\u5165\u8fdb\u884c\u89c4\u8303\u5316<\/strong><\/h2>\n\n\n\n
      \u8bf7\u6ce8\u610f\uff0c\u5728\u6b64\u6b65\u9aa4\u4e4b\u540e\uff0c\u5f62\u72b6(shapes)\u4e0d\u4f1a\u6539\u53d8\uff0c\u503c\u53ea\u662f\u5f52\u4e00\u5316<\/p>\n\n\n\n
      \u8981\u8bb0\u4f4f\u7684\u4e8b\u60c5\uff0c\u6211\u4eec\u9700\u8981\u4e00\u4e2anorm_eps\uff08\u6765\u81ea config\uff09\uff0c\u56e0\u4e3a\u6211\u4eec\u4e0d\u60f3\u610f\u5916\u5730\u5c06 rms \u8bbe\u7f6e\u4e3a 0 \u5e76\u9664\u4ee5 0<\/p>\n\n\n\n
      \u516c\u5f0f\u5982\u4e0b\uff1a<\/p>\n\n\n\n
      \"\"<\/figure>\n\n\n\n
      # def rms_norm(tensor, norm_weights):\n#     rms = (tensor.pow(2).mean(-1, keepdim=True) + norm_eps)**0.5\n#     return tensor * (norm_weights \/ rms)\ndef rms_norm(tensor, norm_weights):\n    return (tensor * torch.rsqrt(tensor.pow(2).mean(-1, keepdim=True) + norm_eps)) * norm_weights<\/pre><\/div>\n\n\n\n
      \u8fd9\u4e2aPython\u51fd\u6570rms_norm<\/code>\u662f\u4e00\u4e2a\u5e94\u7528\u6839\u5747\u65b9\uff08RMS\uff09\u6807\u51c6\u5316\u6280\u672f\u6765\u6807\u51c6\u5316\u5f20\u91cf\u7684\u51fd\u6570\uff0c\u5e76\u4f7f\u7528\u6743\u91cd\u8fdb\u884c\u7f29\u653e\u3002<\/p>\n\n\n\n
      \u51fd\u6570rms_norm<\/code>\u63a5\u53d7\u4e24\u4e2a\u53c2\u6570\uff1a<\/p>\n\n\n\n