{"id":1437,"date":"2025-11-27T23:44:27","date_gmt":"2025-11-27T15:44:27","guid":{"rendered":"https:\/\/lingbo.online\/?p=1437"},"modified":"2025-11-28T00:26:11","modified_gmt":"2025-11-27T16:26:11","slug":"kv-cache","status":"publish","type":"post","link":"https:\/\/lingbo.online\/index.php\/algorithm_learning\/kv-cache\/","title":{"rendered":"KV Cache\u539f\u7406"},"content":{"rendered":"<h3>\u7b80\u4ecb<\/h3>\n<p>KV Cache\u672c\u8d28\u4e0a\u662f\u4e00\u79cd\u7f13\u5b58\u673a\u5236\uff0c\u4e3b\u8981\u5e94\u7528\u5728Transformer\u67b6\u6784\u7684\u6a21\u578b\u4e2d\uff0c\u5c24\u5176\u662f\u751f\u6210\u5f0f\u4efb\u52a1\u7684\u63a8\u7406\u9636\u6bb5\u3002\u5728Transformer\u4e2d\uff0c\u6ce8\u610f\u529b\u673a\u5236\u901a\u8fc7\u8ba1\u7b97Query\u3001key\u548cvalue\u6765\u786e\u5b9a\u8f93\u5165\u5e8f\u5217\u4e2d\u4e0d\u540c\u4f4d\u7f6e\u7684\u5173\u8054\u7a0b\u5ea6\u3002KV Cache\u7684\u4f5c\u7528\u5c31\u662f\u5b58\u50a8\u5df2\u7ecf\u8ba1\u7b97\u8fc7\u7684K\u77e9\u9635\u548cV\u77e9\u9635\uff0c\u907f\u514d\u5176\u5728\u5904\u7406\u65b0\u7684\u8f93\u5165\u65f6\u91cd\u590d\u8ba1\u7b97\uff0c\u4ece\u800c\u63d0\u9ad8\u63a8\u7406\u6548\u7387\u3002<\/p>\n<h3>KV Cache\u539f\u7406<\/h3>\n<p>\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0c\u5f53\u6a21\u578b\u5904\u7406\u7b2c\u4e00\u4e2atoken\u65f6\uff0c\u6b63\u5e38\u8ba1\u7b97K\u77e9\u9635\u548cV\u77e9\u9635\uff0c\u5c06\u5176\u5b58\u50a8\u5230KV Cache\u4e2d\u3002\u5f53\u5904\u7406\u540e\u7eed\u7684token\u65f6\uff0c\u4ec5\u8ba1\u7b97\u65b0\u7684Query\u77e9\u9635\uff0c\u7136\u540e\u8fdb\u884c\u6ce8\u610f\u529b\u8ba1\u7b97\u3002\u5047\u8bbe\u8f93\u5165\u5e8f\u5217\u957f\u5ea6\u4e3a<span class=\"katex-eq\" data-katex-display=\"false\">n<\/span>\uff0c\u7b2c\u4e00\u6b21\u8ba1\u7b97K\u548cV\u65f6\uff0c\u8ba1\u7b97\u91cf\u548c\u5e8f\u5217\u957f\u5ea6\u76f8\u5173\uff0c\u4e3a<span class=\"katex-eq\" data-katex-display=\"false\">O\uff08n\uff09<\/span>\u3002\u5728\u6ca1\u6709KV Cache\u65f6\uff0c\u5982\u679c\u9700\u8981\u751f\u6210\u4e00\u4e2a\u957f\u5ea6\u4e3a<span class=\"katex-eq\" data-katex-display=\"false\">n<\/span>\u7684\u6587\u672c\u5e8f\u5217\uff0c\u6bcf\u751f\u6210\u4e00\u4e2atoken\uff0c\u90fd\u9700\u8981\u5bf9\u6574\u4e2a\u5e8f\u5217\u91cd\u65b0\u8ba1\u7b97K\u548cV\uff0c\u8ba1\u7b97\u590d\u6742\u5ea6\u4e3a<span class=\"katex-eq\" data-katex-display=\"false\">Q\uff08n^2\uff09<\/span>\u3002\u52a0\u5165KV Cache\u540e\uff0c\u540e\u7eed\u751f\u6210\u6bcf\u4e2atoken\u65f6\uff0c\u53ea\u9700\u8981\u8ba1\u7b97\u65b0\u7684Query\uff08\u6bcf\u4e00\u6b65\u7684 Query \u5411\u91cf \u53ea\u7531\u300c\u5f53\u524d\u521a\u8fdb\u6765\u7684\u90a3\u4e00\u4e2a token\u300d\uff08\u52a0\u4e0a\u5b83\u6240\u5728\u7684\u4f4d\u7f6e\u4fe1\u606f\uff09\u4ea7\u751f\uff09\uff0c\u8ba1\u7b97\u91cf\u4e5f\u662f<span class=\"katex-eq\" data-katex-display=\"false\">O\uff08n\uff09<\/span>\u3002\u4f7f\u7528KV Cache\u540e\uff0c\u6574\u4f53\u8ba1\u7b97\u590d\u6742\u5ea6\u4e3a<span class=\"katex-eq\" data-katex-display=\"false\">O\uff08n\uff09<\/span>\uff0c\u6548\u7387\u660e\u663e\u63d0\u9ad8\u3002<\/p>\n<h3>KV Cache\u7684\u4f18\u52bf<\/h3>\n<ul>\n<li>\u52a0\u901f\u63a8\u7406\uff1a\u901a\u8fc7\u907f\u514d\u91cd\u590d\u8ba1\u7b97K\u548cV\uff0cKV Cache\u80fd\u591f\u663e\u8457\u51cf\u5c11\u63a8\u7406\u65f6\u95f4\u3002\u5bf9\u4e8e\u5b9e\u65f6\u6027\u8f83\u9ad8\u7684\u5e94\u7528\uff08\u5982\u804a\u5929\u673a\u5668\u4eba\u7b49\uff09\u975e\u5e38\u91cd\u8981\uff0c\u80fd\u591f\u8ba9\u6a21\u578b\u66f4\u5feb\u5730\u54cd\u5e94\u7528\u6237\u8f93\u5165\u3002<\/li>\n<li>\u8282\u7701\u5185\u5b58\uff1a\u7531\u4e8e\u4e0d\u9700\u8981\u6bcf\u6b21\u63a8\u7406\u65f6\u91cd\u65b0\u8ba1\u7b97\u548c\u5b58\u50a8K\u548cV\uff0cKV Cache\u80fd\u591f\u51cf\u5c11\u5185\u5b58\u4f7f\u7528\uff0c\u8fd9\u4f7f\u5f97\u5728\u8d44\u6e90\u6709\u9650\u7684\u8bbe\u5907\uff08\u5982\u79fb\u52a8\u8bbe\u5907\uff09\u4e0a\u4e5f\u80fd\u591f\u8fd0\u884c\u8f83\u5927\u89c4\u6a21\u7684\u6a21\u578b\u3002<\/li>\n<\/ul>\n<h3>KV Cache\u7684\u5b9e\u73b0<\/h3>\n<h4>\u6570\u636e\u7ed3\u6784<\/h4>\n<p><strong>Key Cache\uff1a<\/strong>\u901a\u5e38\u662f\u4e00\u4e2a\u5f62\u72b6\u4e3a<code>\uff08batch_size,num_heads,seq_len,k_dim\uff09<\/code>\u7684\u5f20\u91cf\uff0c\u7528\u4e8e\u5b58\u50a8Key\u77e9\u9635\u3002\u5176\u4e2d\uff0c<code>batch_size<\/code>\u8868\u793a\u4e00\u6b21\u5904\u7406\u7684\u6837\u672c\u6570\u91cf\uff0c<code>num_heads<\/code>\u662f\u6ce8\u610f\u529b\u5934\u7684\u6570\u91cf\uff0c<code>seq_len<\/code>\u662f\u5e8f\u5217\u957f\u5ea6\uff0c<code>k_dim<\/code>\u662fKey\u7684\u7ef4\u5ea6\u3002<br \/>\n<strong>Value Cache\uff1a<\/strong>\u5f62\u72b6\u4e3a<code>\uff08batch_size,num_heads,seq_len,v_dim\uff09<\/code>\u7684\u5f20\u91cf\uff0c\u7528\u4e8e\u5b58\u50a8Value\u77e9\u9635\u3002\u5176\u4e2d<code>v_dim<\/code>\u662fValue\u7684\u7ef4\u5ea6\u3002<\/p>\n<h4>\u8ba1\u7b97\u6d41\u7a0b<\/h4>\n<ul>\n<li>\u521d\u59cb\u5316\u9636\u6bb5\uff1a\u8f93\u5165\u521d\u59cbtoken\uff0c\u7ecf\u8fc7\u7ebf\u6027\u53d8\u6362\u5f97\u5230Q\u3001K\u3001V\u3002\u5c06K\u548cV\u5b58\u50a8\u5230KV Cache\u4e2d\uff0c\u4e3a\u540e\u7eed\u7ed3\u7b97\u51c6\u5907\u597d\u201c\u539f\u6750\u6599\u201d\u3002<\/li>\n<li>\u540e\u7eedtoken\u5904\u7406\u9636\u6bb5\uff1a\u8f93\u5165\u65b0\u7684token\uff0c\u8ba1\u7b97\u65b0\u7684query\uff0c\u5e76\u4eceKV Cache\u4e2d\u8bfb\u53d6\u4e4b\u524d\u5b58\u50a8\u7684K\u548cV\uff0c\u8fdb\u884c\u6ce8\u610f\u529b\u8ba1\u7b97\uff0c\u5f97\u5230\u8f93\u51fa\u3002<br \/>\n\u5c06\u65b0\u751f\u6210\u7684K\u548cV\u8ffd\u52a0\u5230KV Cache\u4e2d\uff0c\u4ee5\u4fbf\u540e\u7eed\u4f7f\u7528\u3002\u968f\u7740\u65b0token\u7684\u4e0d\u65ad\u8f93\u5165\uff0cKV Cache\u5c06\u4e0d\u65ad\u66f4\u65b0\u548c\u6269\u5145\uff0c\u4fdd\u8bc1\u6a21\u578b\u80fd\u591f\u59cb\u7ec8\u8fdb\u884c\u9ad8\u6548\u7684\u63a8\u7406\u3002<\/li>\n<\/ul>\n<h4>\u4ee3\u7801\u793a\u4f8b<\/h4>\n<pre><code class=\"language-python\">import torch\nimport torch.nn as nn\nimport math\n\nclass CachedAttention(nn.Module):\n    def __init__(self, d_model, num_heads):\n        super().__init__()\n        self.d_model = d_model\n        self.num_heads = num_heads\n        self.head_dim = d_model \/\/ num_heads\n        # \u5b9a\u4e49\u7ebf\u6027\u53d8\u6362\u5c42\uff0c\u5c06\u8f93\u5165\u6620\u5c04\u5230Query\u3001Key\u548cValue\u7a7a\u95f4\n        self.q_proj = nn.Linear(d_model, d_model)\n        self.k_proj = nn.Linear(d_model, d_model)\n        self.v_proj = nn.Linear(d_model, d_model)\n        # \u5b9a\u4e49\u8f93\u51fa\u7ebf\u6027\u53d8\u6362\u5c42\uff0c\u5c06\u6ce8\u610f\u529b\u8ba1\u7b97\u7ed3\u679c\u6620\u5c04\u56de\u539f\u7ef4\u5ea6\n        self.out_proj = nn.Linear(d_model, d_model)\n\n    def forward(self, x, kv_cache=None):\n        b, t, c = x.shape\n        # \u5c06\u8f93\u5165x\u901a\u8fc7\u7ebf\u6027\u53d8\u6362\u5f97\u5230Query\uff0c\u5e76\u8c03\u6574\u5f62\u72b6\u548c\u7ef4\u5ea6\n        q = self.q_proj(x).view(b, t, self.num_heads, self.head_dim).transpose(1, 2)\n        # \u5c06\u8f93\u5165x\u901a\u8fc7\u7ebf\u6027\u53d8\u6362\u5f97\u5230Key\uff0c\u5e76\u8c03\u6574\u5f62\u72b6\u548c\u7ef4\u5ea6\n        k = self.k_proj(x).view(b, t, self.num_heads, self.head_dim).transpose(1, 2)\n        # \u5c06\u8f93\u5165x\u901a\u8fc7\u7ebf\u6027\u53d8\u6362\u5f97\u5230Value\uff0c\u5e76\u8c03\u6574\u5f62\u72b6\u548c\u7ef4\u5ea6\n        v = self.v_proj(x).view(b, t, self.num_heads, self.head_dim).transpose(1, 2)\n\n        if kv_cache is not None:\n            cached_k, cached_v = kv_cache\n            # \u5c06\u7f13\u5b58\u4e2d\u7684Key\u548c\u5f53\u524d\u8ba1\u7b97\u7684Key\u62fc\u63a5\u8d77\u6765\n            k = torch.cat((cached_k, k), dim=2)\n            # \u5c06\u7f13\u5b58\u4e2d\u7684Value\u548c\u5f53\u524d\u8ba1\u7b97\u7684Value\u62fc\u63a5\u8d77\u6765\n            v = torch.cat((cached_v, v), dim=2)\n\n        # \u8ba1\u7b97\u6ce8\u610f\u529b\u5206\u6570\uff0c\u8fd9\u91cc\u9664\u4ee5\u6839\u53f7\u4e0bhead_dim\u662f\u4e3a\u4e86\u7f29\u653e\n        attn = (q @ k.transpose(-2, -1)) * (1.0 \/ math.sqrt(self.head_dim))\n        # \u5bf9\u6ce8\u610f\u529b\u5206\u6570\u8fdb\u884csoftmax\u5f52\u4e00\u5316\n        attn = attn.softmax(dim=-1)\n        # \u6839\u636e\u6ce8\u610f\u529b\u5206\u6570\u5bf9Value\u8fdb\u884c\u52a0\u6743\u6c42\u548c\n        y = (attn @ v).transpose(1, 2).contiguous().view(b, t, c)\n        # \u901a\u8fc7\u8f93\u51fa\u7ebf\u6027\u53d8\u6362\u5c42\u5f97\u5230\u6700\u7ec8\u8f93\u51fa\n        y = self.out_proj(y)\n        return y, (k, v)<\/code><\/pre>\n<h3>\u5e94\u7528\u8981\u70b9<\/h3>\n<p>\u5728\u5b9e\u9645\u5de5\u7a0b\u5e94\u7528\u4e2d\uff0cKV Cache \u8fd8\u9762\u4e34\u4e00\u4e9b\u6311\u6218\u548c\u9700\u8981\u4f18\u5316\u7684\u5730\u65b9\u3002\u4f8b\u5982\uff0c\u7f13\u5b58\u5927\u5c0f\u7684\u52a8\u6001\u8c03\u6574\u662f\u4e00\u4e2a\u5173\u952e\u95ee\u9898\u3002\u968f\u7740\u8f93\u5165\u5e8f\u5217\u957f\u5ea6\u7684\u4e0d\u65ad\u589e\u52a0\uff0c\u7f13\u5b58\u5360\u7528\u7684\u5185\u5b58\u4e5f\u4f1a\u9010\u6e10\u589e\u5927\uff0c\u5982\u679c\u4e0d\u52a0\u4ee5\u63a7\u5236\uff0c\u53ef\u80fd\u4f1a\u5bfc\u81f4\u5185\u5b58\u6ea2\u51fa\u3002\u56e0\u6b64\uff0c\u9700\u8981\u6839\u636e\u5b9e\u9645\u60c5\u51b5\u52a8\u6001\u8c03\u6574\u7f13\u5b58\u5927\u5c0f\uff0c\u6bd4\u5982\u5f53\u7f13\u5b58\u8fbe\u5230\u4e00\u5b9a\u9608\u503c\u65f6\uff0c\u5bf9\u7f13\u5b58\u8fdb\u884c\u538b\u7f29\u6216\u8005\u820d\u5f03\u90e8\u5206\u8f83\u65e9\u7684\u7f13\u5b58\u6570\u636e\u3002\u5728\u591a GPU \u73af\u5883\u4e0b\uff0cKV Cache \u7684\u7ba1\u7406\u4e5f\u53d8\u5f97\u66f4\u52a0\u590d\u6742\u3002\u4e0d\u540c GPU \u4e4b\u95f4\u9700\u8981\u8fdb\u884c\u6709\u6548\u7684\u6570\u636e\u540c\u6b65\uff0c\u786e\u4fdd\u6bcf\u4e2a GPU \u90fd\u80fd\u83b7\u53d6\u5230\u6b63\u786e\u7684\u7f13\u5b58\u6570\u636e\uff0c\u540c\u65f6\u8fd8\u8981\u907f\u514d\u6570\u636e\u4f20\u8f93\u5e26\u6765\u7684\u989d\u5916\u5f00\u9500\u3002\u53ef\u4ee5\u91c7\u7528\u5206\u5e03\u5f0f\u7f13\u5b58\u7ba1\u7406\u7b56\u7565\uff0c\u5408\u7406\u5206\u914d\u7f13\u5b58\u6570\u636e\u5230\u5404\u4e2a GPU\uff0c\u63d0\u9ad8\u6574\u4f53\u7684\u8ba1\u7b97\u6548\u7387\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7b80\u4ecb KV Cache\u672c\u8d28\u4e0a\u662f\u4e00\u79cd\u7f13\u5b58\u673a\u5236\uff0c\u4e3b\u8981\u5e94\u7528\u5728Transformer\u67b6\u6784\u7684\u6a21\u578b\u4e2d\uff0c\u5c24\u5176\u662f\u751f\u6210\u5f0f\u4efb\u52a1\u7684\u63a8\u7406\u9636\u6bb5\u3002\u5728Trans &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"emotion":"","emotion_color":"","title_style":"","license":"","footnotes":""},"categories":[45,2],"tags":[72,70,71],"class_list":["post-1437","post","type-post","status-publish","format-standard","hentry","category-llms","category-algorithm_learning","tag-attention","tag-kv-cache","tag-transformer"],"_links":{"self":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1437","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/comments?post=1437"}],"version-history":[{"count":4,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1437\/revisions"}],"predecessor-version":[{"id":1441,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1437\/revisions\/1441"}],"wp:attachment":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/media?parent=1437"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/categories?post=1437"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/tags?post=1437"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}