
{"id":11344,"date":"2026-08-24T05:02:38","date_gmt":"2026-08-23T21:02:38","guid":{"rendered":"https:\/\/infernews.com\/blog\/?page_id=11344"},"modified":"2026-09-24T20:52:35","modified_gmt":"2026-09-24T12:52:35","slug":"deepseek","status":"publish","type":"page","link":"https:\/\/infernews.com\/blog\/deepseek\/","title":{"rendered":"DeepSeek"},"content":{"rendered":"\n<table id=\"tablepress-11\" class=\"tablepress tablepress-id-11\">\n<thead>\n<tr class=\"row-1\">\n\t<th class=\"column-1\">\u9805\u76ee<\/th><th class=\"column-2\">\u7db2\u5740<\/th>\n<\/tr>\n<\/thead>\n<tbody class=\"row-striping row-hover\">\n<tr class=\"row-2\">\n\t<td class=\"column-1\">Awesome DeepSeek Harness<\/td><td class=\"column-2\"><a rel=\"noopener\" target=\"_blank\" href=\"https:\/\/github.com\/0xsline\/awesome-deepseek-harness\">https:\/\/github.com\/0xsline\/awesome-deepseek-harness<\/a><\/td>\n<\/tr>\n<tr class=\"row-3\">\n\t<td class=\"column-1\">DeepSeek Harness Ultimate<\/td><td class=\"column-2\"><a rel=\"noopener\" href=\"https:\/\/github.com\/18126295767-cell\/deepseek-harness-ultimate\" target=\"_blank\">https:\/\/github.com\/18126295767-cell\/deepseek-harness-ultimate<\/a><\/td>\n<\/tr>\n<tr class=\"row-4\">\n\t<td class=\"column-1\"><\/td><td class=\"column-2\"><\/td>\n<\/tr>\n<tr class=\"row-5\">\n\t<td class=\"column-1\"><\/td><td class=\"column-2\"><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<!-- #tablepress-11 from cache -->\n\n\n\n<p class=\"wp-block-paragraph\">DeepSeek V4.1 Flash<\/p>\n\n\n\n<style>\nhtml {\n  scroll-behavior: smooth;\n}\n#wp-scoped-article {\n  font-family: -apple-system, BlinkMacSystemFont, \"Segoe UI\", Roboto, \"Noto Sans TC\", sans-serif;\n  color: #1e293b;\n  line-height: 1.8;\n  font-size: 22px; \n  margin: 1.5rem 0;\n  box-sizing: border-box;\n  -webkit-font-smoothing: antialiased;\n}\n#wp-scoped-article * {\n  box-sizing: border-box;\n}\n#wp-scoped-article h1 {\n  font-size: 1.8rem;\n  font-weight: 800;\n  color: #0f172a;\n  line-height: 1.35;\n  margin-bottom: 1.25rem;\n}\n#wp-scoped-article h2 {\n  font-size: 1.5rem;\n  font-weight: 700;\n  color: #0f172a;\n  margin-top: 2.25rem;\n  margin-bottom: 1rem;\n  padding-bottom: 0.5rem;\n  border-bottom: 2px solid #e2e8f0;\n  display: flex;\n  align-items: center;\n  gap: 0.5rem;\n  scroll-margin-top: 80px; \n}\n#wp-scoped-article h3 {\n  font-size: 1.25rem;\n  font-weight: 600;\n  color: #1e293b;\n  margin-top: 1.6rem;\n  margin-bottom: 0.75rem;\n  scroll-margin-top: 80px; \n}\n#wp-scoped-article p {\n  margin-bottom: 1.2rem;\n  color: #334155;\n}\n#wp-scoped-article a {\n  color: #2563eb;\n  text-decoration: none;\n}\n#wp-scoped-article a:hover {\n  text-decoration: underline;\n}\n#wp-scoped-article .in-toc {\n  background: #f8fafc;\n  border: 1px solid #e2e8f0;\n  border-radius: 8px;\n  padding: 1.5rem;\n  margin: 1.5rem 0;\n}\n#wp-scoped-article .in-toc a {\n  text-decoration: none;\n  color: #2563eb;\n}\n#wp-scoped-article .in-toc a:hover {\n  text-decoration: underline;\n}\n#wp-scoped-article .in-toc ul {\n  list-style: none;\n  padding-left: 0;\n  margin: 0;\n}\n#wp-scoped-article .in-toc li {\n  margin-bottom: 0.5rem;\n}\n#wp-scoped-article .in-tag-group {\n  display: flex;\n  flex-wrap: wrap;\n  gap: 6px;\n  margin-bottom: 1.25rem;\n}\n#wp-scoped-article .in-tag {\n  display: inline-block;\n  padding: 2px 10px;\n  border-radius: 9999px;\n  font-size: 17px;\n  font-weight: 600;\n  background: #eff6ff;\n  color: #1d4ed8;\n  border: 1px solid #bfdbfe;\n}\n#wp-scoped-article .in-callout-danger {\n  background-color: #fef2f2;\n  border-left: 5px solid #ef4444;\n  border-radius: 0 8px 8px 0;\n  padding: 1.15rem 1.25rem;\n  margin: 1.5rem 0;\n  font-size: 1rem;\n}\n#wp-scoped-article .in-callout-info {\n  background-color: #eff6ff;\n  border-left: 5px solid #3b82f6;\n  border-radius: 0 8px 8px 0;\n  padding: 1.15rem 1.25rem;\n  margin: 1.5rem 0;\n  font-size: 1rem;\n}\n#wp-scoped-article .in-badge {\n  display: inline-flex;\n  align-items: center;\n  justify-content: center;\n  width: 28px;\n  height: 28px;\n  border-radius: 6px;\n  background: #2563eb;\n  color: #ffffff;\n  font-size: 18px;\n  font-weight: 700;\n  margin-right: 6px;\n  flex-shrink: 0;\n}\n#wp-scoped-article .in-terminal-box {\n  background: #0d1117;\n  border: 1px solid #30363d;\n  border-radius: 8px;\n  margin: 1.5rem 0;\n  overflow: hidden;\n  box-shadow: 0 4px 12px rgba(0, 0, 0, 0.12);\n}\n#wp-scoped-article .in-terminal-header {\n  background: #161b22;\n  padding: 8px 14px;\n  display: flex;\n  align-items: center;\n  justify-content: space-between;\n  border-bottom: 1px solid #30363d;\n  font-family: ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, monospace;\n  font-size: 14px;\n  color: #8b949e;\n}\n#wp-scoped-article .in-terminal-dots {\n  display: flex;\n  gap: 6px;\n}\n#wp-scoped-article .in-dot {\n  width: 10px;\n  height: 10px;\n  border-radius: 50%;\n  display: inline-block;\n}\n#wp-scoped-article .in-dot-red { background: #ff5f56; }\n#wp-scoped-article .in-dot-yellow { background: #ffbd2e; }\n#wp-scoped-article .in-dot-green { background: #27c93f; }\n#wp-scoped-article pre.in-code {\n  margin: 0;\n  padding: 16px;\n  background: #0d1117;\n  color: #c9d1d9;\n  font-family: ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, monospace;\n  font-size: 15.5px;\n  line-height: 1.6;\n  overflow-x: auto;\n  white-space: pre;\n  border: none;\n}\n#wp-scoped-article .in-table-wrapper {\n  overflow-x: auto;\n  margin: 1.5rem 0;\n  border: 1px solid #e2e8f0;\n  border-radius: 8px;\n}\n#wp-scoped-article table.in-table {\n  width: 100%;\n  border-collapse: collapse;\n  font-size: 20px; \n  text-align: left;\n  margin: 0;\n}\n#wp-scoped-article table.in-table th {\n  background-color: #f8fafc;\n  color: #334155;\n  font-weight: 600;\n  padding: 12px 16px;\n  border-bottom: 1px solid #e2e8f0;\n  font-size: 20px; \n}\n#wp-scoped-article table.in-table td {\n  padding: 12px 16px;\n  border-bottom: 1px solid #f1f5f9;\n  color: #475569;\n  font-size: 20px; \n}\n#wp-scoped-article .in-checklist {\n  display: grid;\n  grid-template-columns: repeat(auto-fit, minmax(240px, 1fr));\n  gap: 10px;\n  margin: 1.5rem 0;\n}\n#wp-scoped-article .in-checklist-item {\n  display: flex;\n  align-items: center;\n  gap: 10px;\n  padding: 10px 14px;\n  background: #f8fafc;\n  border: 1px solid #e2e8f0;\n  border-radius: 8px;\n  font-size: 20px;\n}\n#wp-scoped-article .in-check-icon {\n  color: #10b981;\n  font-weight: bold;\n}\n<\/style>\n\n<div id=\"wp-scoped-article\">\n  <div class=\"in-tag-group\">\n    <span class=\"in-tag\">\u5927\u578b\u8a9e\u8a00\u6a21\u578b<\/span>\n    <span class=\"in-tag\">DeepSeek V4.1 Flash<\/span>\n    <span class=\"in-tag\">\u63a8\u8ad6\u52a0\u901f<\/span>\n    <span class=\"in-tag\">CED \u67b6\u69cb<\/span>\n    <span class=\"in-tag\">\u5feb\u53d6\u512a\u5316<\/span>\n  <\/div>\n\n  <h1>DeepSeek V4.1 Flash \u6a21\u578b\u67b6\u69cb\u8207\u63a8\u8ad6\u512a\u5316\u6280\u8853\u898f\u683c\u5256\u6790<\/h1>\n\n  <p>DeepSeek V4.1 Flash \u91dd\u5c0d\u73fe\u4ee3\u5927\u8a9e\u8a00\u6a21\u578b\u63a8\u8ad6\u4e2d\u6700\u6602\u8cb4\u7684\u968e\u6bb5\u9032\u884c\u4e86\u6df1\u5c64\u5e95\u5c64\u9769\u65b0\u3002\u900f\u904e\u5168\u65b0\u7684\u56e0\u679c\u7de8\u78bc\u89e3\u78bc\u5668\uff08CED\uff09\u67b6\u69cb\u8207\u8de8\u5c64\u5feb\u53d6\u91cd\u7528\u6280\u8853\uff0c\u6210\u529f\u964d\u4f4e\u4e86\u9ad8\u9054\u4e00\u534a\u7684\u9810\u586b\u5145\u904b\u7b97\u8ca0\u64d4\uff0c\u4e26\u5927\u5e45\u7bc0\u7701\u6bcf Token \u6240\u4f54\u7528\u7684 KV \u5feb\u53d6\u8a18\u61b6\u9ad4\u7a7a\u9593\u3002<\/p>\n\n  <div class=\"in-toc\">\n    <strong>\u6587\u7ae0\u76ee\u9304<\/strong>\n    <ul>\n      <li><a href=\"#section-ced\">1. \u56e0\u679c\u7de8\u78bc\u89e3\u78bc\u5668\uff08CED\uff09\u8207\u904b\u7b97\u5206\u6d41\u6a5f\u5236<\/a><\/li>\n      <li><a href=\"#section-kv\">2. \u7dda\u6027 KV \u5feb\u53d6\u6295\u5f71\u8207\u6ed1\u52d5\u7a97\u53e3\u91cd\u64ad<\/a><\/li>\n      <li><a href=\"#section-memory\">3. \u8a18\u61b6\u9ad4\u6700\u4f73\u5316\uff1aEngram \u6a21\u7d44\u8207 CSA 2<\/a><\/li>\n      <li><a href=\"#section-comparison\">4. \u6838\u5fc3\u67b6\u69cb\u6280\u8853\u898f\u683c\u5b8c\u6574\u5c0d\u7167\u8868<\/a><\/li>\n      <li><a href=\"#section-checklist\">5. \u63a8\u8ad6\u4f3a\u670d\u5668\u90e8\u7f72\u786c\u9ad4\u6aa2\u6838\u6e05\u55ae<\/a><\/li>\n    <\/ul>\n  <\/div>\n\n  <h2 id=\"section-ced\"><span class=\"in-badge\">1<\/span> \u56e0\u679c\u7de8\u78bc\u89e3\u78bc\u5668\uff08CED\uff09\u8207\u904b\u7b97\u5206\u6d41\u6a5f\u5236<\/h2>\n  <p>\u50b3\u7d71\u7684 Transformer \u67b6\u69cb\u5728\u8655\u7406 Prompt \u6642\uff0c\u5fc5\u9808\u5c07\u6240\u6709\u8a5e\u5143\u901a\u904e\u5b8c\u6574\u4e14\u6602\u8cb4\u7684\u89e3\u78bc\u5c64\u3002DeepSeek V4.1 Flash \u63a1\u7528\u56e0\u679c\u7de8\u78bc\u89e3\u78bc\u5668\uff08Causal Encoder Decoder, CED\uff09\uff0c\u5c07\u6a21\u578b\u5206\u70ba\u5c0d\u7a31\u7684\u7de8\u78bc\u5668\u8207\u89e3\u78bc\u5668\uff1a<\/p>\n\n  <div class=\"in-callout-info\">\n    <strong>\u6838\u5fc3\u67b6\u69cb\u512a\u52e2\uff1a<\/strong>\u4ee5 40 \u5c64\u7684\u6a21\u578b\u70ba\u4f8b\uff0c\u9810\u586b\u5145\u968e\u6bb5\u50c5\u9700\u6fc0\u6d3b\u524d\u534a\u6bb5 L\/2 \u5c64\uff08\u5373\u7b2c 1 \u81f3 20 \u5c64\uff09\u7684\u7de8\u78bc\u5668\u3002\u5728\u6b64\u968e\u6bb5\uff0c\u6bcf Token \u50c5\u9700\u8abf\u7528\u7d04 8B \u6d3b\u8e8d\u53c3\u6578\uff0c\u76f8\u6bd4\u89e3\u78bc\u968e\u6bb5\u7684 16B \u6d3b\u8e8d\u53c3\u6578\uff0c\u76f4\u63a5\u7701\u4e0b\u8fd1 50% \u7684\u8a08\u7b97\u6210\u672c\u3002\n  <\/div>\n\n  <p>\u9019\u9805\u7a81\u7834\u4f7f\u5f97\u9ad8\u4f75\u767c\u63a8\u8ad6\u4f3a\u670d\u5668\u5728\u9762\u5c0d\u6975\u9577\u4e0a\u4e0b\u6587\u6642\uff0c\u80fd\u5c07\u6642\u9593\u8207\u8a18\u61b6\u9ad4\u74f6\u9838\u6709\u6548\u79fb\u51fa\u9810\u586b\u5145\u968e\u6bb5\u3002<\/p>\n\n  <h2 id=\"section-kv\"><span class=\"in-badge\">2<\/span> \u7dda\u6027 KV \u5feb\u53d6\u6295\u5f71\u8207\u6ed1\u52d5\u7a97\u53e3\u91cd\u64ad<\/h2>\n  <p>\u70ba\u4e86\u4f7f\u89e3\u78bc\u5c64\u7121\u9700\u5b8c\u6574\u8a08\u7b97\u524d\u5411\u50b3\u64ad\u5373\u53ef\u7372\u5f97\u4e0a\u4e0b\u6587\u8cc7\u8a0a\uff0c\u7cfb\u7d71\u76f4\u63a5\u81ea\u7de8\u78bc\u5668\u9802\u5c64\u9032\u884c\u591a\u5c64\u6295\u5f71\uff0c\u4e26\u91dd\u5c0d\u908a\u754c\u7a97\u53e3\u5be6\u65bd\u5b89\u5168\u91cd\u64ad\uff1a<\/p>\n\n  <div class=\"in-terminal-box\">\n    <div class=\"in-terminal-header\">\n      <div class=\"in-terminal-dots\">\n        <span class=\"in-dot in-dot-red\"><\/span>\n        <span class=\"in-dot in-dot-yellow\"><\/span>\n        <span class=\"in-dot in-dot-green\"><\/span>\n      <\/div>\n      <span>ced_kv_replay_pipeline.py<\/span>\n    <\/div>\n    <pre class=\"in-code\"># [Step 1] \u81ea\u7de8\u78bc\u5668\u6700\u5f8c\u4e00\u5c64\u9032\u884c\u7dda\u6027\u6295\u5f71\u4ee5\u69cb\u5efa\u5feb\u53d6\nencoder_hidden_states = run_encoder(prompt_tokens)  # \u50c5\u8a08\u7b97 L\/2 \u5c64\nkv_cache = linear_projection_to_decoder(encoder_hidden_states)\n\n# [Step 2] \u91dd\u5c0d\u6ed1\u52d5\u7a97\u53e3 (SWA) \u672b\u7aef 128 Tokens \u57f7\u884c\u5c40\u90e8\u91cd\u64ad\nboundary_tokens = prompt_tokens[-128:]\nswa_reconstructed_cache = decoder_swa_replay(boundary_tokens)\nfinal_kv_cache = merge_cache(kv_cache, swa_reconstructed_cache)\n\n# [Step 3] \u9032\u5165\u5e38\u898f\u89e3\u78bc\u968e\u6bb5 (\u6bcf Token \u6d3b\u8e8d\u53c3\u6578: 16B)\nrun_decoding(final_kv_cache)<\/pre>\n  <\/div>\n\n  <div class=\"in-callout-danger\">\n    <strong>\u77ed\u6587\u672c\u63a8\u8ad6\u6ce8\u610f\u4e8b\u9805\uff1a<\/strong>SWA Replay \u5fc5\u9808\u91dd\u5c0d\u6700\u5f8c 128 \u500b Token \u91cd\u65b0\u8dd1\u904e\u89e3\u78bc\u8def\u5f91\u3002\u82e5\u8f38\u5165 Prompt \u9577\u5ea6\u4f4e\u65bc\u6216\u63a5\u8fd1 128 \u500b Token\uff0c\u9810\u586b\u5145\u7684\u5169\u500d\u52a0\u901f\u512a\u52e2\u5c07\u6703\u88ab\u908a\u754c\u91cd\u65b0\u8a08\u7b97\u7684\u958b\u92b7\u62b5\u6d88\u3002\n  <\/div>\n\n  <h2 id=\"section-memory\"><span class=\"in-badge\">3<\/span> \u8a18\u61b6\u9ad4\u6700\u4f73\u5316\uff1aEngram \u6a21\u7d44\u8207 CSA 2<\/h2>\n  <p>\u9664\u4e86\u904b\u7b97\u91cf\u5206\u6d41\uff0cV4.1 Flash \u5f15\u5165\u4e86 Engram \u6a21\u7d44\u8207\u58d3\u7e2e\u7a00\u758f\u6ce8\u610f\u529b\u7b2c 2 \u7248\uff08CSA 2\uff09\uff1a<\/p>\n  <p><strong>Engram \u7368\u7acb\u8a18\u61b6\u9ad4\u901a\u9053\uff1a<\/strong>\u900f\u904e 196B \u5916\u90e8\u53c3\u6578\u8207 8 \u500b Hash Heads\uff0c\u5728\u7b2c 1 \u5c64\u8207\u7b2c 15 \u5c64\u76f4\u63a5\u6ce8\u5165\u5c40\u90e8 N-gram \u8a18\u61b6\u6a21\u5f0f\uff0c\u5c07\u91cd\u8907\u8a9e\u8a00\u6a23\u5f0f\u7684\u8a18\u61b6\u529f\u80fd\u81ea\u8a9e\u8a00\u9aa8\u5e79\u4e2d\u89e3\u8026\u3002<\/p>\n  <p><strong>Single-pass MHC\uff1a<\/strong>\u5c07\u591a\u6a19\u982d\u63a7\u5236\u7684\u6b98\u5dee\u8a08\u7b97\u5408\u4f75\u70ba\u55ae\u4e00 Kernel\uff0c\u5c07\u7b97\u5b50\u8907\u96dc\u5ea6\u58d3\u81f3\u7406\u8ad6\u4e0b\u754c\uff0c\u6d88\u9664\u6642\u5e8f\u4f9d\u8cf4\u6027\u4e26\u5927\u5e45\u7de9\u89e3 GPU \u9ad8\u983b\u5bec\u8a18\u61b6\u9ad4\uff08HBM\uff09\u7684\u5f80\u8fd4\u8b80\u5beb\u5ef6\u9072\u3002<\/p>\n\n  <h2 id=\"section-comparison\"><span class=\"in-badge\">4<\/span> \u6838\u5fc3\u67b6\u69cb\u6280\u8853\u898f\u683c\u5b8c\u6574\u5c0d\u7167\u8868<\/h2>\n  <p>\u4ee5\u4e0b\u6574\u7406 DeepSeek V4.1 Flash \u5404\u9805\u4e3b\u8981\u6280\u8853\u7d44\u4ef6\u7684\u904b\u4f5c\u76ee\u6a19\u8207\u5177\u9ad4\u898f\u683c\u6578\u64da\uff1a<\/p>\n\n  <div class=\"in-table-wrapper\">\n    <table class=\"in-table\">\n      <thead>\n        <tr>\n          <th>\u6280\u8853\u7d44\u4ef6<\/th>\n          <th>\u6838\u5fc3\u529f\u80fd\u8207\u6a5f\u5236<\/th>\n          <th>\u95dc\u9375\u53c3\u6578\u6216\u6307\u6a19<\/th>\n          <th>\u63a8\u8ad6\u968e\u6bb5\u6548\u80fd\u5f71\u97ff<\/th>\n        <\/tr>\n      <\/thead>\n      <tbody>\n        <tr>\n          <td><strong>\u56e0\u679c\u7de8\u78bc\u89e3\u78bc\u5668 (CED)<\/strong><\/td>\n          <td>\u5c07\u7d50\u69cb\u62c6\u70ba Encoder \u8207 Decoder\uff0c\u524d\u5411\u8a08\u7b97\u76f4\u63a5\u6295\u5f71\u81f3\u5feb\u53d6\u3002<\/td>\n          <td>\u5404 L\/2 \u5c64\uff1b\u9810\u586b\u5145 8B \u6d3b\u8e8d\u53c3\u6578\uff0c\u89e3\u78bc 16B \u6d3b\u8e8d\u53c3\u6578\u3002<\/td>\n          <td>\u9810\u586b\u5145\u968e\u6bb5\u8a08\u7b97\u6210\u672c\u6e1b\u534a\uff0c\u5cf0\u503c\u8a18\u61b6\u9ad4\u5927\u5e45\u964d\u4f4e\u3002<\/td>\n        <\/tr>\n        <tr>\n          <td><strong>KV Cache Projection<\/strong><\/td>\n          <td>\u4ee5\u7dda\u6027\u5c64\u5c07 Encoder \u96b1\u542b\u72c0\u614b\u6295\u5c04\u81f3\u5404\u89e3\u78bc\u5c64\u3002<\/td>\n          <td>Linear Projection \u7368\u7acb\u6b0a\u91cd\u77e9\u9663\u3002<\/td>\n          <td>\u7701\u53bb\u9810\u586b\u5145\u6642\u9010\u5c64\u8a08\u7b97 Decoder \u7684\u8a18\u61b6\u9ad4\u8207\u904b\u7b97\u958b\u92b7\u3002<\/td>\n        <\/tr>\n        <tr>\n          <td><strong>SWA Replay<\/strong><\/td>\n          <td>\u91dd\u5c0d\u672b\u7aef\u8996\u7a97\u91cd\u5efa Decoder \u81ea\u8eab\u5feb\u53d6\u72c0\u614b\u3002<\/td>\n          <td>\u91cd\u65b0\u8a08\u7b97\u6700\u5f8c 128 \u500b Token \u7684\u89e3\u78bc\u8def\u5f91\u3002<\/td>\n          <td>\u9577\u6587\u672c\u7dad\u6301\u8fd1 2 \u500d\u52a0\u901f\uff1b\u6975\u77ed\u6587\u672c\u52a0\u901f\u5e45\u5ea6\u6709\u9650\u3002<\/td>\n        <\/tr>\n        <tr>\n          <td><strong>Engram \u6a21\u7d44<\/strong><\/td>\n          <td>\u5c07\u91cd\u8907\u8a9e\u8a00\u6a21\u5f0f\u89e3\u8026\u81f3\u7368\u7acb\u8a18\u61b6\u9ad4\u69fd\u3002<\/td>\n          <td>196B \u984d\u5916\u53c3\u6578\u3001256 \u7dad\u5ea6\u30018 Hash Heads\u3002<\/td>\n          <td>\u6e1b\u5c11\u9aa8\u5e79\u7db2\u8def\u53cd\u8986\u8a08\u7b97\u5c40\u90e8\u6a23\u5f0f\u7684\u8cc7\u6e90\u8ca0\u64d4\u3002<\/td>\n        <\/tr>\n        <tr>\n          <td><strong>CSA 2<\/strong><\/td>\n          <td>\u8de8\u5c64 KV \u72c0\u614b\u9078\u64c7\u6027\u91cd\u7528\u6a5f\u5236\u3002<\/td>\n          <td>\u652f\u63f4 Full\u3001Index\u3001Reuse \u4e09\u7a2e\u914d\u7f6e\u6a21\u5f0f\u3002<\/td>\n          <td>\u6bcf Token \u5feb\u53d6\u8a18\u61b6\u9ad4\u4f54\u7528\u964d\u81f3\u524d\u4ee3\u6a21\u578b\u7684\u56db\u5206\u4e4b\u4e00\u3002<\/td>\n        <\/tr>\n        <tr>\n          <td><strong>Single-pass MHC<\/strong><\/td>\n          <td>\u6b98\u5dee\u7db2\u8def\u4ee3\u6578\u512a\u5316\u8207\u904b\u7b97\u5143\u878d\u5408\u3002<\/td>\n          <td>3 \u500b Kernel \u5408\u4f75\u70ba 1 \u500b\uff1b\u964d\u81f3\u7406\u8ad6\u4e0b\u754c (2n+2d)\u3002<\/td>\n          <td>\u6d88\u9664\u8a18\u61b6\u9ad4\u983b\u5bec\u74f6\u9838\uff0c\u986f\u8457\u62c9\u5347 GPU \u541e\u5410\u91cf\u3002<\/td>\n        <\/tr>\n      <\/tbody>\n    <\/table>\n  <\/div>\n\n  <h2 id=\"section-checklist\"><span class=\"in-badge\">5<\/span> \u63a8\u8ad6\u4f3a\u670d\u5668\u90e8\u7f72\u786c\u9ad4\u6aa2\u6838\u6e05\u55ae<\/h2>\n  <p>\u5728\u63a8\u8ad6\u53e2\u96c6\u4e0a\u67b6\u8a2d V4.1 Flash \u524d\uff0c\u8acb\u78ba\u8a8d\u4f3a\u670d\u5668\u786c\u9ad4\u8207\u9a45\u52d5\u5df2\u7b26\u5408\u4ee5\u4e0b\u5fc5\u8981\u689d\u4ef6\uff1a<\/p>\n\n  <div class=\"in-checklist\">\n    <div class=\"in-checklist-item\">\n      <span class=\"in-check-icon\">\u2713<\/span>\n      <span>\u652f\u63f4 FP8 \u6df7\u5408\u7cbe\u5ea6\u77e9\u9663\u904b\u7b97<\/span>\n    <\/div>\n    <div class=\"in-checklist-item\">\n      <span class=\"in-check-icon\">\u2713<\/span>\n      <span>GPU HBM \u983b\u5bec\u9ad8\u65bc 2.0 TB\/s<\/span>\n    <\/div>\n    <div class=\"in-checklist-item\">\n      <span class=\"in-check-icon\">\u2713<\/span>\n      <span>\u63a8\u8ad6\u6846\u67b6\u652f\u63f4\u55ae\u4e00 Kernel \u7b97\u5b50\u878d\u5408<\/span>\n    <\/div>\n    <div class=\"in-checklist-item\">\n      <span class=\"in-check-icon\">\u2713<\/span>\n      <span>\u9a45\u52d5\u7a0b\u5f0f\u652f\u63f4 SWA \u908a\u754c\u91cd\u64ad\u6392\u7a0b<\/span>\n    <\/div>\n  <\/div>\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>DeepSeek V4.1 Flash \u5927\u578b\u8a9e\u8a00\u6a21\u578b DeepSeek V4.1 Flash \u63a8\u8ad6\u52a0\u901f CED \u67b6\u69cb \u5feb\u53d6\u512a\u5316 DeepSeek V4.1 Flash \u6a21\u578b\u67b6\u69cb\u8207\u63a8\u8ad6\u512a\u5316\u6280\u8853\u898f\u683c\u5256\u6790 DeepSeek V4.1 Flash \u91dd\u5c0d\u73fe\u4ee3\u5927\u8a9e\u8a00\u6a21\u578b\u63a8\u8ad6\u4e2d\u6700\u6602\u8cb4\u7684\u968e\u6bb5\u9032\u884c\u4e86\u6df1\u5c64\u5e95\u5c64\u9769\u65b0\u3002\u900f\u904e\u5168\u65b0\u7684\u56e0\u679c\u7de8\u78bc\u89e3\u78bc\u5668\uff08CED\uff09\u67b6\u69cb\u8207\u8de8\u5c64\u5feb\u53d6\u91cd\u7528\u6280\u8853\uff0c\u6210\u529f\u964d\u4f4e\u4e86\u9ad8\u9054\u4e00\u534a\u7684\u9810\u586b\u5145\u904b\u7b97\u8ca0\u64d4\uff0c\u4e26\u5927\u5e45\u7bc0\u7701\u6bcf Token \u6240\u4f54\u7528\u7684 KV \u5feb\u53d6\u8a18\u61b6\u9ad4\u7a7a\u9593\u3002 \u6587\u7ae0\u76ee\u9304 1. \u56e0\u679c\u7de8\u78bc\u89e3\u78bc\u5668\uff08CED\uff09\u8207\u904b\u7b97\u5206\u6d41\u6a5f\u5236 2. \u7dda\u6027 KV \u5feb\u53d6\u6295\u5f71\u8207\u6ed1\u52d5\u7a97\u53e3\u91cd\u64ad 3. \u8a18\u61b6\u9ad4\u6700\u4f73\u5316\uff1aEngram \u6a21\u7d44\u8207 CSA 2 4. \u6838\u5fc3\u67b6\u69cb\u6280\u8853\u898f\u683c\u5b8c\u6574\u5c0d\u7167\u8868 5. \u63a8\u8ad6\u4f3a\u670d\u5668\u90e8\u7f72\u786c\u9ad4\u6aa2\u6838\u6e05\u55ae 1 \u56e0\u679c\u7de8\u78bc\u89e3\u78bc\u5668\uff08CED\uff09\u8207\u904b\u7b97\u5206\u6d41\u6a5f\u5236 \u50b3\u7d71\u7684 Transformer \u67b6\u69cb\u5728\u8655\u7406 Prompt \u6642\uff0c\u5fc5\u9808\u5c07\u6240\u6709\u8a5e\u5143\u901a\u904e\u5b8c\u6574\u4e14\u6602\u8cb4\u7684\u89e3\u78bc\u5c64\u3002DeepSeek V4.1 Flash \u63a1\u7528\u56e0\u679c\u7de8\u78bc\u89e3\u78bc\u5668\uff08Causal Encoder Decoder, CED\uff09\uff0c\u5c07\u6a21\u578b\u5206\u70ba\u5c0d\u7a31\u7684\u7de8\u78bc\u5668\u8207\u89e3\u78bc\u5668\uff1a \u6838\u5fc3\u67b6\u69cb\u512a\u52e2\uff1a\u4ee5 40 \u5c64\u7684\u6a21\u578b\u70ba\u4f8b\uff0c\u9810\u586b\u5145\u968e\u6bb5\u50c5\u9700\u6fc0\u6d3b\u524d\u534a\u6bb5 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"parent":0,"menu_order":0,"comment_status":"closed","ping_status":"closed","template":"","meta":{"wpai_generated_summary":"","wpai_meta_description":"","footnotes":""},"class_list":["post-11344","page","type-page","status-publish","hentry"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/pages\/11344","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/page"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=11344"}],"version-history":[{"count":2,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/pages\/11344\/revisions"}],"predecessor-version":[{"id":12607,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/pages\/11344\/revisions\/12607"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=11344"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}