{"id":12654,"date":"2026-09-27T23:10:16","date_gmt":"2026-09-27T15:10:16","guid":{"rendered":"https:\/\/infernews.com\/blog\/?page_id=12654"},"modified":"2026-09-28T00:09:20","modified_gmt":"2026-09-27T16:09:20","slug":"transformer-llm-vs-dits-cost","status":"publish","type":"page","link":"https:\/\/infernews.com\/blog\/transformer-llm-vs-dits-cost\/","title":{"rendered":"Transformer LLM \u8207 DiTs \u8ab0\u7684\u6210\u672c\u66f4\u6602\u8cb4\uff1f"},"content":{"rendered":"\n<figure class=\"wp-block-image size-full\"><img data-dominant-color=\"d8e0e5\" data-has-transparency=\"false\" style=\"--dominant-color: #d8e0e5;\" loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"572\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/09\/image-18.png\" alt=\"\" class=\"wp-image-12664 not-transparent\" srcset=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/09\/image-18.png 1024w, https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/09\/image-18-300x168.png 300w, https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/09\/image-18-768x429.png 768w\" \/><\/figure>\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\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: none;\n  opacity: 0.8;\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 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\">\u751f\u6210\u5f0f AI \u67b6\u69cb<\/span>\n    <span class=\"in-tag\">Transformer LLM<\/span>\n    <span class=\"in-tag\">DiT \u64f4\u6563\u6a21\u578b<\/span>\n    <span class=\"in-tag\">RTX 4090 \u6548\u80fd\u76e3\u63a7<\/span>\n    <span class=\"in-tag\">\u786c\u9ad4\u74f6\u9838\u5256\u6790<\/span>\n  <\/div>\n\n  <h1>\u751f\u6210 1 \u500b Token \u7684\u4ee3\u50f9\uff1aLLM \u8207 DiT \u7b97\u529b\u74f6\u9838\u5256\u6790\u3001RTX 4090 \u5be6\u8e10\u8207\u95dc\u9375\u76e3\u63a7\u6307\u6a19\u6821\u6b63<\/h1>\n\n  <p>\u5728\u7576\u524d\u751f\u6210\u5f0f AI \u67b6\u69cb\u4e2d\uff0c\u81ea\u56de\u6b78\u5927\u578b\u8a9e\u8a00\u6a21\u578b\uff08Autoregressive LLM\uff09\u8207\u64f4\u6563 Transformer\uff08DiT\uff09\u96d6\u7136\u5e95\u5c64\u7686\u57fa\u65bc\u6ce8\u610f\u529b\u6a5f\u5236\uff08Self-Attention\uff09\uff0c\u4f46\u5728\u300c\u7522\u751f 1 \u500b Token\u300d\u7684\u7269\u7406\u610f\u6db5\u3001\u786c\u9ad4\u8cc7\u6e90\u74f6\u9838\u8207\u7522\u51fa\u4ee3\u50f9\u4e0a\u5b58\u5728\u672c\u8cea\u5dee\u7570\u3002\u672c\u6587\u9664\u4e86\u6df1\u5165\u5256\u6790\u5169\u8005\u7684\u67b6\u69cb\u958b\u92b7\uff0c\u66f4\u63d0\u4f9b\u4e00\u5957\u53ef\u5728\u6d88\u8cbb\u7d1a\u65d7\u8266\u986f\u793a\u5361\uff08NVIDIA GeForce RTX 4090\uff09\u4e0a\u7a69\u5b9a\u904b\u884c\u7684\u5be6\u9ad4\u76e3\u63a7\u8173\u672c\uff0c\u4e26\u7279\u5225\u91dd\u5c0d\u696d\u754c\u6700\u5e38\u8aa4\u8b80\u7684 <code>utilization.memory<\/code> \u6838\u5fc3\u6307\u6a19\u9032\u884c\u6df1\u5ea6\u91d0\u6e05\u3002<\/p>\n\n  <div class=\"in-toc\">\n    <strong>\u76ee\u9304\u5c0e\u89bd<\/strong>\n    <ul>\n      <li><a href=\"#token-concept\">1. \u6982\u5ff5\u89e3\u69cb\uff1a\u96e2\u6563\u6587\u5b57 Token vs \u7a7a\u9593 Patch Token<\/a><\/li>\n      <li><a href=\"#bottleneck-analysis\">2. \u786c\u9ad4\u67b6\u69cb\u74f6\u9838\uff1a\u8a18\u61b6\u9ad4\u983b\u5bec vs \u7b97\u529b\u5bc6\u96c6<\/a><\/li>\n      <li><a href=\"#cost-comparison\">3. \u6210\u672c\u5c0d\u6c7a\uff1a\u55ae Token \u5230\u53ef\u7528\u8f38\u51fa\u7684\u7e3d\u91cf\u5206\u6790<\/a><\/li>\n      <li><a href=\"#consumer-gpu-script\">4. RTX 4090 \u5be6\u9ad4\u76e3\u63a7\u8173\u672c\u8207\u6307\u6a19\u5b9a\u7fa9\u6821\u6b63<\/a><\/li>\n      <li><a href=\"#rtx4090-caveats\">5. \u6d88\u8cbb\u7d1a\u74b0\u5883\u95dc\u9375\u6ce8\u610f\u4e8b\u9805\u8207\u6eab\u63a7\u9677\u9631<\/a><\/li>\n    <\/ul>\n  <\/div>\n\n  <div class=\"in-callout-info\">\n    <strong>\u6838\u5fc3\u5148\u9a57\u7d50\u8ad6\uff1a<\/strong> \u55ae\u8ad6\u300c\u55ae\u6b21\u524d\u5411\u50b3\u64ad\uff08Single Forward Pass\uff09\u8655\u7406 1 \u500b Token\u300d\uff0cLLM \u8207 DiT \u6d88\u8017\u7684\u7406\u8ad6\u904b\u7b97\u91cf\u5728\u5bc6\u96c6\u77e9\u9663\u4e58\u6cd5\u7c21\u5316\u4f30\u7b97\u4e0b\u7686\u7d04\u70ba\u6a21\u578b\u53c3\u6578\u91cf\u4e4b\u5169\u500d\uff08$2 \\times P$\uff09\u3002\u7136\u800c\u5728\u5b8c\u6210 1 \u500b\u6700\u7d42\u53ef\u7528\u7684\u8cc7\u8a0a\u55ae\u5143\u6642\uff0cDiT \u5fc5\u9808\u7d93\u904e\u6578\u5341\u500b\u53bb\u566a\u6b65\u9a5f\u53cd\u8986\u91cd\u7b97\u6574\u500b\u7a7a\u9593 Token \u5e8f\u5217\uff0c\u4e14\u5e38\u898b\u7684\u6587\u5b57\u751f\u6210\u5f71\u50cf\u4efb\u52d9\u4f34\u96a8\u7121\u5206\u985e\u5668\u5f15\u5c0e\uff08CFG\uff09\u5c07\u4f7f\u8a08\u7b97\u91cf\u518d\u6b21\u7ffb\u500d\uff0c\u52a0\u4e0a\u7aef\u5230\u7aef\u4e0d\u53ef\u6216\u7f3a\u7684 VAE \u89e3\u78bc\u958b\u92b7\uff0c\u5176\u6574\u9ad4\u786c\u9ad4\u7b97\u529b\u4ee3\u50f9\u9060\u9ad8\u65bc\u7d14\u81ea\u56de\u6b78 LLM\u3002\n  <\/div>\n\n  <h2 id=\"token-concept\"><span class=\"in-badge\">1<\/span>\u6982\u5ff5\u89e3\u69cb\uff1a\u96e2\u6563\u6587\u5b57 Token vs \u7a7a\u9593 Patch Token<\/h2>\n  <p>\u5728\u8a55\u4f30\u904b\u7b97\u6210\u672c\u524d\uff0c\u5fc5\u9808\u5148\u5340\u5206\u5169\u8005\u5c0d\u65bc\u300cToken\u300d\u5b9a\u7fa9\u4e0a\u7684\u6839\u672c\u6b67\u7570\uff1a<\/p>\n  <p><strong>\u81ea\u56de\u6b78\u8a9e\u8a00\u6a21\u578b\uff08Transformer LLM\uff09\uff1a<\/strong> \u63a1\u7528\u4e00\u7dad\u96e2\u6563\u6587\u5b57\u5207\u5206\uff08\u5982 BPE\u3001WordPiece\uff09\u3002\u751f\u6210 1 \u500b Token \u4ee3\u8868\u7531\u6a21\u578b\u5728\u8a5e\u5f59\u5eab\uff08Vocabulary\uff09\u4e2d\u9810\u6e2c\u4e0b\u4e00\u500b\u55ae\u5b57\u7684\u6a5f\u7387\u5206\u4f48\u3002\u6b64 Token \u4e00\u65e6\u5b8c\u6210\u63a1\u6a23\u4fbf\u6c38\u4e45\u56fa\u5b9a\uff0c\u4e26\u76f4\u63a5\u5beb\u5165 KV \u5feb\u53d6\uff08KV Cache\uff09\u4f9b\u5f8c\u7e8c\u905e\u63a8\u4f7f\u7528\uff0c\u4e0d\u9700\u8981\u518d\u6b21\u91cd\u65b0\u8a08\u7b97\u3002<\/p>\n  <p><strong>\u64f4\u6563 Transformer\uff08DiT\uff09\uff1a<\/strong> \u501f\u9451 Vision Transformer\uff08ViT\uff09\u908f\u8f2f\uff0c\u5c07 VAE \u7de8\u78bc\u5f8c\u7684\u9023\u7e8c\u6f5b\u5728\u7279\u5fb5\u5716\uff08Latent Feature Map\uff09\u5207\u5206\u6210\u591a\u500b\u56fa\u5b9a\u5c3a\u5bf8\u7684\u7a7a\u9593\u5340\u584a\uff08Patches\uff0c\u4f8b\u5982 2\u00d72 \u6216 4\u00d74\uff09\u3002\u6bcf\u500b Patch \u5373\u70ba 1 \u500b\u8996\u89ba Token\u3002\u4f46\u4e0d\u540c\u65bc\u6587\u5b57\u7684\u662f\uff0c\u8a72\u8996\u89ba Token \u4e26\u975e\u4e00\u6b21\u751f\u6210\u5b8c\u7562\uff0c\u800c\u662f\u5728\u9023\u7e8c\u591a\u6b65\uff08\u5982 20\uff5e50 \u500b Timesteps\uff09\u53bb\u566a\u904e\u7a0b\u4e2d\uff0c\u8207\u5168\u57df Token \u5171\u540c\u5728\u96d9\u5411\u6ce8\u610f\u529b\u6a5f\u5236\u4e0b\u6301\u7e8c\u88ab\u66f4\u65b0\u8207\u8986\u5beb\u3002\u82e5\u555f\u7528\u6587\u751f\u5716\u6216\u6587\u751f\u5f71\u7247\u5e38\u898b\u7684\u7121\u5206\u985e\u5668\u5f15\u5c0e\uff08Classifier-Free Guidance, CFG\uff09\uff0c\u6bcf\u4e00\u6b65\u53bb\u566a\u9084\u9700\u540c\u6642\u57f7\u884c\u689d\u4ef6\uff08Conditional\uff09\u8207\u975e\u689d\u4ef6\uff08Unconditional\uff09\u5169\u6b21\u5b8c\u6574\u524d\u5411\u50b3\u64ad\uff0c\u4f7f\u8a08\u7b97\u958b\u92b7\u76f4\u63a5\u7ffb\u500d\u3002\u6b64\u5916\uff0c\u5728\u6240\u6709\u53bb\u566a\u6b65\u9a5f\u5b8c\u6210\u5f8c\uff0c\u9084\u5fc5\u9808\u7d93\u904e VAE \u89e3\u78bc\u5668\uff08VAE Decoder\uff09\u5c07\u6f5b\u5728\u7279\u5fb5\u9084\u539f\u70ba\u50cf\u7d20\u5f71\u50cf\uff0c\u5e36\u4f86\u984d\u5916\u7684\u986f\u5b58\u5cf0\u503c\u8207\u904b\u7b97\u5ef6\u9072\u3002<\/p>\n\n  <h2 id=\"compute-bottleneck\"><span class=\"in-badge\">2<\/span>\u786c\u9ad4\u67b6\u69cb\u74f6\u9838\uff1a\u8a18\u61b6\u9ad4\u983b\u5bec vs \u7b97\u529b\u5bc6\u96c6<\/h2>\n  <p>\u4e0d\u540c\u7684\u751f\u6210\u6a21\u5f0f\u6c7a\u5b9a\u4e86\u5169\u8005\u5728 GPU \u904b\u7b97\u55ae\u5143\u4e0a\u7684\u8cc7\u6e90\u53d7\u9650\u7dad\u5ea6\uff1a<\/p>\n\n  <div class=\"in-checklist\">\n    <div class=\"in-checklist-item\">\n      <span class=\"in-check-icon\">\u2713<\/span>\n      <span><strong>LLM \u89e3\u78bc\u968e\u6bb5\uff08\u55ae\u4e26\u767c Batch Size = 1\uff09\uff1a<\/strong>\u6975\u5ea6\u53d7\u9650\u65bc\u8a18\u61b6\u9ad4\u983b\u5bec\uff08Memory Bandwidth Bound\uff09<\/span>\n    <\/div>\n    <div class=\"in-checklist-item\">\n      <span class=\"in-check-icon\">\u2713<\/span>\n      <span><strong>DiT \u53bb\u566a\u968e\u6bb5\uff1a<\/strong>\u9ad8\u5ea6\u53d7\u9650\u65bc\u5f35\u91cf\u904b\u7b97\u55ae\u5143\u7b97\u529b\uff08Compute-Bound\uff09<\/span>\n    <\/div>\n    <div class=\"in-checklist-item\">\n      <span class=\"in-check-icon\">\u2713<\/span>\n      <span><strong>KV \u5feb\u53d6\u8ca0\u64d4\uff1a<\/strong>LLM \u96a8\u8457\u4e0a\u4e0b\u6587\u9577\u5ea6\u7dda\u6027\u64f4\u5f35<\/span>\n    <\/div>\n    <div class=\"in-checklist-item\">\n      <span class=\"in-check-icon\">\u2713<\/span>\n      <span><strong>\u5168\u57df\u96d9\u5411\u8a08\u7b97\uff1a<\/strong>DiT \u6bcf\u4e00\u6b65\u7dad\u6301\u56fa\u5b9a Patch \u5e8f\u5217\u4e26\u884c\u8a08\u7b97<\/span>\n    <\/div>\n  <\/div>\n\n  <p>\u503c\u5f97\u5f37\u8abf\u7684\u662f\uff0c<strong>\u5728\u55ae\u4e26\u767c\u6216\u55ae\u8acb\u6c42\uff08Batch Size = 1\uff09\u7684\u60c5\u5883\u4e0b\uff0cLLM \u89e3\u78bc\u6975\u5ea6\u53d7\u9650\u65bc\u8a18\u61b6\u9ad4\u983b\u5bec<\/strong>\uff0c\u56e0\u70ba GPU \u5fc5\u9808\u5c07\u6578\u5341\u5104\u751a\u81f3\u767e\u5104\u898f\u6a21\u7684\u5b8c\u6574\u6b0a\u91cd\u77e9\u9663\u5f9e HBM\/\u986f\u5b58\u8f09\u5165\u81f3\u9ad8\u901f\u5feb\u53d6\uff0c\u537b\u50c5\u70ba\u9019\u55ae\u4e00 Token \u8a08\u7b97\u9577\u5ea6\u70ba 1 \u7684\u67e5\u8a62\u5411\u91cf\uff0c\u5c0e\u81f4\u786c\u9ad4\u7b97\u529b\u5229\u7528\u7387\uff08MFU\uff09\u901a\u5e38\u4f4e\u65bc 30%\u3002\u53ea\u6709\u900f\u904e\u9023\u7e8c\u6279\u6b21\u8655\u7406\uff08Continuous Batching\uff09\u5c07 Batch Size \u653e\u5927\uff0cLLM \u624d\u80fd\u9010\u6b65\u8f49\u79fb\u70ba\u904b\u7b97\u5bc6\u96c6\u578b\uff1b\u76f8\u5c0d\u5730\uff0cDiT \u5929\u7136\u5177\u5099\u6210\u767e\u4e0a\u5343\u500b\u7a7a\u9593 Patch Token \u540c\u6b65\u4e26\u884c\uff0c\u5373\u4f7f\u5728\u55ae\u5f35\u5f71\u50cf\u751f\u6210\u6642\u4e5f\u80fd\u8f15\u9b06\u69a8\u4e7e\u5f35\u91cf\u6838\u5fc3\u7b97\u529b\u3002<\/p>\n\n  <h2 id=\"cost-comparison\"><span class=\"in-badge\">3<\/span>\u6210\u672c\u5c0d\u6c7a\uff1a\u55ae Token \u5230\u53ef\u7528\u8f38\u51fa\u7684\u7e3d\u91cf\u5206\u6790<\/h2>\n  <p>\u6211\u5011\u5c07\u81ea\u56de\u6b78 LLM \u8207 DiT \u5728\u63a8\u8ad6\u67b6\u69cb\u8207\u6210\u672c\u7dad\u5ea6\u9032\u884c\u6a6b\u5411\u5c0d\u6bd4\uff1a<\/p>\n\n  <div class=\"in-table-wrapper\">\n    <table class=\"in-table\">\n      <thead>\n        <tr>\n          <th>\u8a55\u4f30\u7dad\u5ea6<\/th>\n          <th>Transformer LLM (\u81ea\u56de\u6b78\u89e3\u78bc)<\/th>\n          <th>Diffusion Transformer (DiT \u53bb\u566a)<\/th>\n        <\/tr>\n      <\/thead>\n      <tbody>\n        <tr>\n          <td><strong>\u751f\u6210\u6a5f\u5236<\/strong><\/td>\n          <td>\u56e0\u679c\u55ae\u5411\uff08Causal Auto-regressive\uff09\u9010\u5b57\u905e\u63a8<\/td>\n          <td>\u96d9\u5411\u5168\u6ce8\u610f\u529b\uff08Bidirectional\uff09\u591a\u6b65\u53cd\u8986\u8fed\u4ee3\u53bb\u566a<\/td>\n        <\/tr>\n        <tr>\n          <td><strong>Token \u751f\u547d\u9031\u671f<\/strong><\/td>\n          <td>\u4e00\u6b21\u751f\u6210\uff0c\u6c38\u4e45\u56fa\u5b9a\uff0c\u5feb\u53d6\u81f3 KV Cache<\/td>\n          <td>\u8de8\u6b65\u9a5f\u8fed\u4ee3\u66f4\u65b0\uff08\u901a\u5e38\u6b77\u7d93 20\uff5e50 \u6b21\u91cd\u7b97\uff09<\/td>\n        <\/tr>\n        <tr>\n          <td><strong>\u786c\u9ad4\u8cc7\u6e90\u74f6\u9838<\/strong><\/td>\n          <td><strong>\u55ae\u8acb\u6c42\uff08Batch = 1\uff09\u4e0b\u6975\u5ea6\u53d7\u9650\u65bc\u8a18\u61b6\u9ad4\u983b\u5bec<\/strong><\/td>\n          <td><strong>\u904b\u7b97\u55ae\u5143\u7b97\u529b\uff08FLOPs \/ Compute Bound\uff09<\/strong><\/td>\n        <\/tr>\n        <tr>\n          <td><strong>\u55ae Token \u524d\u5411 FLOPs<\/strong><\/td>\n          <td>\u7d04 $2 \\times P$\uff08*\u8a3b\uff1a\u5bc6\u96c6\u77e9\u9663\u4e58\u6cd5\u4f30\u7b97\uff09<\/td>\n          <td>\u7d04 $2 \\times P$\uff08*\u8a3b\uff1a\u5bc6\u96c6\u77e9\u9663\u4e58\u6cd5\u4f30\u7b97\uff09<\/td>\n        <\/tr>\n        <tr>\n          <td><strong>CFG \u984d\u5916\u500d\u7387<\/strong><\/td>\n          <td>\u7121\uff08\u539f\u751f\u652f\u63f4\u689d\u4ef6\u6587\u5b57\u751f\u6210\uff09<\/td>\n          <td><strong>\u901a\u5e38\u555f\u7528 CFG\uff0c\u55ae\u6b65\u524d\u5411\u6b21\u6578 $\\times 2$<\/strong><\/td>\n        <\/tr>\n        <tr>\n          <td><strong>\u786c\u9ad4\u5229\u7528\u7387 (MFU)<\/strong><\/td>\n          <td>\u55ae\u8acb\u6c42\u8f03\u4f4e\uff1b\u4f9d\u8cf4\u9ad8\u6279\u6b21\uff08Large Batching\uff09\u63d0\u5347<\/td>\n          <td>\u55ae\u8acb\u6c42\u5373\u5177\u5099\u9ad8\u4e26\u884c\u5ea6\uff0cGPU \u5229\u7528\u7387\u8f03\u9ad8<\/td>\n        <\/tr>\n        <tr>\n          <td><strong>\u7aef\u5230\u7aef\u4ea4\u4ed8\u4ee3\u50f9<\/strong><\/td>\n          <td>\u6309\u751f\u6210 Token \u6578\u91cf\u8a08\u8cbb\uff0c\u55ae Token \u7d55\u5c0d\u6210\u672c\u6975\u4f4e<\/td>\n          <td>\u55ae\u5e45\u5f71\u50cf\u7d2f\u7a4d\u6578\u5343 Token $\\times$ \u591a\u6b65\u9a5f $\\times$ CFG\uff0c\u52a0\u4e0a <strong>VAE Decode<\/strong> \u986f\u5b58\u8207\u7b97\u529b\u958b\u92b7\uff0c\u7e3d\u6210\u672c\u6975\u9ad8<\/td>\n        <\/tr>\n      <\/tbody>\n    <\/table>\n  <\/div>\n\n  <div class=\"in-callout-info\">\n    <strong>*\u516c\u5f0f\u7cbe\u78ba\u5ea6\u8a3b\u89e3\uff1a<\/strong> \u4e0a\u8868\u6240\u5217\u7684 $2 \\times P$ \u70ba\u5ffd\u7565 Attention \u4ea4\u4e92\u9805\u7684\u5bc6\u96c6\u77e9\u9663\u4e58\u6cd5\uff08Linear GEMM\uff09\u7c21\u5316\u4f30\u7b97\u3002\u5728\u8d85\u9577\u4e0a\u4e0b\u6587\uff08LLM \u9700\u9762\u5c0d\u8d85\u5927 KV Cache \u7684\u77e9\u9663\u5411\u91cf\u76f8\u4e58\uff09\u6216\u9ad8\u89e3\u6790\u5ea6\u7a7a\u9593 Patch \u5e8f\u5217\uff08DiT \u9700\u57f7\u884c $O(N^2)$ \u7684\u5168\u57df\u96d9\u5411 Attention\uff09\u60c5\u5883\u4e0b\uff0c\u5fc5\u9808\u984d\u5916\u8a08\u5165\u6ce8\u610f\u529b\u77e9\u9663\u8207 Softmax \u4ea4\u4e92\u9805\u7684\u8a08\u7b97\u91cf\u3002\n  <\/div>\n\n  <h2 id=\"consumer-gpu-script\"><span class=\"in-badge\">4<\/span>RTX 4090 \u5be6\u9ad4\u76e3\u63a7\u8173\u672c\u8207\u6307\u6a19\u5b9a\u7fa9\u6821\u6b63<\/h2>\n  <p>\u70ba\u4e86\u5728\u4f3a\u670d\u5668\u7aef\u6216\u5de5\u4f5c\u7ad9\u7cbe\u6e96\u6355\u6349\u63a8\u8ad6\u6642\u7684\u771f\u5be6\u786c\u9ad4\u958b\u92b7\uff0c\u5de5\u7a0b\u5e2b\u5e38\u8abf\u7528 Shell \u6307\u4ee4\u78bc\u9032\u884c\u63a1\u6a23\u3002\u91dd\u5c0d\u6d88\u8cbb\u7d1a\u65d7\u8266 <strong>NVIDIA GeForce RTX 4090<\/strong>\uff0c\u96d6\u7136\u7f3a\u4e4f\u8cc7\u6599\u4e2d\u5fc3\u7d1a\u7684 ECC \u932f\u8aa4\u7d71\u8a08\u8207 DCGM \u5b88\u8b77\u884c\u7a0b\uff0c\u4f46\u900f\u904e NVML \u7684\u57fa\u790e\u67e5\u8a62\uff0c\u4f9d\u7136\u80fd\u5b8c\u6574\u64f7\u53d6\u6838\u5fc3\u7b97\u529b\u3001\u8a18\u61b6\u9ad4\u3001\u529f\u8017\u8207\u6563\u71b1\u6307\u6a19\u3002<\/p>\n\n  <div class=\"in-callout-danger\">\n    <strong>\u91cd\u5927\u6982\u5ff5\u66f4\u6b63\uff1autilization.memory \u7684\u771f\u5be6\u542b\u7fa9<\/strong><br>\n    \u8a31\u591a\u5de5\u7a0b\u5e2b\u5bb9\u6613\u5c07 <code>utilization.memory<\/code> \u8aa4\u89e3\u70ba\u300c\u986f\u793a\u5361\u8a18\u61b6\u9ad4\u5df2\u4f7f\u7528\u7684\u5bb9\u91cf\u767e\u5206\u6bd4\uff08VRAM Usage %\uff09\u300d\uff0c<strong>\u9019\u662f\u4e00\u500b\u975e\u5e38\u56b4\u91cd\u7684\u8aa4\u89e3\uff01<\/strong><br><br>\n    \u6839\u64da NVIDIA NVML \u5b98\u65b9\u6280\u8853\u898f\u7bc4\uff1a<br>\n    \u2022 <strong><code>utilization.memory<\/code>\uff08\u8a18\u61b6\u9ad4\u63a7\u5236\u5668\u5229\u7528\u7387\uff09\uff1a<\/strong> \u4ee3\u8868\u5728\u904e\u53bb\u7684\u53d6\u6a23\u9031\u671f\u5167\uff0cGPU \u8a18\u61b6\u9ad4\u63a7\u5236\u5668\uff08Memory Controller\uff09\u8655\u65bc\u8b80\u53d6\u6216\u5beb\u5165\u6d3b\u8e8d\u72c0\u614b\u7684\u6642\u9593\u767e\u5206\u6bd4\u3002\u5b83\u53cd\u6620\u7684\u662f<strong>\u8a18\u61b6\u9ad4\u5b58\u53d6\u983b\u5bec\u7684\u5fd9\u788c\u7a0b\u5ea6<\/strong>\uff0c\u800c\u975e\u5bb9\u91cf\u3002<br>\n    \u2022 <strong><code>memory.used \/ memory.total<\/code>\uff08\u986f\u5b58\u7a7a\u9593\u4f7f\u7528\u7387\uff09\uff1a<\/strong> \u82e5\u8981\u8a55\u4f30\u300c\u76ee\u524d 24GB \u986f\u5b58\u88ab\u6a21\u578b\u6b0a\u91cd\u3001KV \u5feb\u53d6\u6216\u53bb\u566a\u7279\u5fb5\u4f54\u7528\u4e86\u591a\u5c11\u7a7a\u9593\u300d\uff0c\u5fc5\u9808\u7531 <code>memory.used<\/code> \u6578\u503c\u9664\u4ee5 <code>memory.total<\/code> \u63db\u7b97\u5f97\u51fa\u3002\n  <\/div>\n\n  <p>\u4ee5\u4e0b\u70ba\u91dd\u5c0d RTX 4090 \u6700\u4f73\u5316\u7de8\u5beb\u7684\u5be6\u9ad4\u76e3\u63a7\u8173\u672c\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>rtx4090_inference_monitor.sh<\/span>\n    <\/div>\n    <pre class=\"in-code\">#!\/usr\/bin\/env bash\n# ==============================================================================\n# NVIDIA RTX 4090 \u63a8\u8ad6\u74b0\u5883\u5c08\u7528\u6548\u80fd\u8207\u71b1\u80fd\u76e3\u63a7\u8173\u672c\n# \u9069\u7528\u5c08\u6848\uff1avLLM, Ollama, ComfyUI, TensorRT-LLM \u4f3a\u670d\u5668\n# ==============================================================================\n\nset -euo pipefail\n\nLOG_FILE=\"rtx4090_profile_$(date +%Y%m%d_%H%M%S).csv\"\nINTERVAL=1\n\necho \"[\u8cc7\u8a0a] \u6b63\u5728\u521d\u59cb\u5316 RTX 4090 \u76e3\u63a7\u7ba1\u7dda...\"\necho \"[\u8cc7\u8a0a] \u53d6\u6a23\u6578\u64da\u5c07\u5373\u6642\u5beb\u5165: ${LOG_FILE}\"\n\n# \u5beb\u5165 CSV \u6a19\u984c\u5217\uff08\u7279\u5225\u5340\u5206\u8a18\u61b6\u9ad4\u63a7\u5236\u5668\u983b\u5bec\u8207\u986f\u5b58\u5bb9\u91cf\uff09\n# mem_ctrl_busy_pct = utilization.memory\uff08\u8a18\u61b6\u9ad4\u532f\u6d41\u6392\u5fd9\u788c\u767e\u5206\u6bd4\uff09\n# mem_used_mb       = \u5be6\u969b\u986f\u5b58\u4f54\u7528\u5bb9\u91cf (MB)\necho \"timestamp,gpu_util_pct,mem_ctrl_busy_pct,mem_used_mb,mem_free_mb,power_draw_w,temp_c,fan_speed_pct,sm_clock_mhz\" > \"${LOG_FILE}\"\n\ntrap 'echo -e \"\\n[\u5b8c\u6210] \u76e3\u63a7\u884c\u7a0b\u5df2\u4e2d\u6b62\uff0c\u65e5\u8a8c\u5b8c\u6574\u5132\u5b58\u3002\"; exit 0' INT TERM\n\nwhile true; do\n    TIMESTAMP=$(date -u +\"%Y-%m-%dT%H:%M:%SZ\")\n    \n    # \u900f\u904e nvidia-smi \u64f7\u53d6\u5c08\u5c6c\u6307\u6a19\uff08\u6392\u9664\u4f01\u696d\u7d1a\u7121\u6548\u6b04\u4f4d\uff09\n    STATS=$(nvidia-smi --query-gpu=utilization.gpu,utilization.memory,memory.used,memory.free,power.draw,temperature.gpu,fan.speed,clocks.current.sm \\\n        --format=csv,noheader,nounits | tr -d ' ')\n    \n    echo \"${TIMESTAMP},${STATS}\" >> \"${LOG_FILE}\"\n    sleep \"${INTERVAL}\"\ndone<\/pre>\n  <\/div>\n\n  <h2 id=\"rtx4090-caveats\"><span class=\"in-badge\">5<\/span>\u6d88\u8cbb\u7d1a\u74b0\u5883\u95dc\u9375\u6ce8\u610f\u4e8b\u9805\u8207\u6eab\u63a7\u9677\u9631<\/h2>\n  <p>\u5728 RTX 4090 \u4e0a\u5be6\u6e2c LLM \u8207 DiT \u63a8\u8ad6\u6642\uff0c\u9700\u7279\u5225\u9632\u7bc4\u786c\u9ad4\u908a\u754c\u5c0d\u5be6\u9a57\u6578\u64da\u7684\u5e72\u64fe\uff1a<\/p>\n  <p><strong>1. \u6563\u71b1\u964d\u983b\uff08Thermal Throttling\uff09\uff1a<\/strong> \u6d88\u8cbb\u7d1a\u986f\u793a\u5361\u6563\u71b1\u591a\u4f9d\u8cf4\u98a8\u51b7\u6563\u71b1\u7247\u3002\u5728\u9577\u6642\u9593\u904b\u7b97 DiT \u7b49\u8a08\u7b97\u5bc6\u96c6\u578b\u4efb\u52d9\u6642\uff0c\u6838\u5fc3\u6eab\u5ea6\u82e5\u7a81\u7834 83\u00b0C\uff0c\u986f\u793a\u5361\u6703\u4e3b\u52d5\u58d3\u4f4e\u4e32\u6d41\u591a\u8655\u7406\u5668\u6642\u8108\uff08SM Clock\uff09\uff0c\u5c0e\u81f4\u63a8\u8ad6\u8017\u6642\u88ab\u52d5\u62c9\u9577\uff0c\u7522\u751f\u975e\u7dda\u6027\u7684\u6548\u80fd\u8aa4\u5dee\u3002<\/p>\n  <p><strong>2. 24GB \u986f\u5b58\u786c\u9650\u5236\uff08OOM \u5d29\u6f70\u98a8\u96aa\uff09\uff1a<\/strong> \u76f8\u8f03\u65bc\u8cc7\u6599\u4e2d\u5fc3\u7d1a\u52a0\u901f\u5361\uff08\u5982 A100 80GB\uff09\uff0cRTX 4090 \u7684\u986f\u5b58\u76f8\u7576\u7dca\u6e4a\u3002\u5728\u81ea\u56de\u6b78 LLM \u55ae\u8acb\u6c42\u9577\u4e0a\u4e0b\u6587\u6216\u9ad8\u4e26\u884c\u63a8\u7406\u6642\uff0c\u9700\u56b4\u5bc6\u76e3\u63a7 <code>mem_used_mb<\/code>\uff0c\u9632\u7bc4 KV \u5feb\u53d6\u788e\u7247\u5316\u6ea2\u4f4d\uff1b\u800c\u5728\u57f7\u884c\u9ad8\u89e3\u6790\u5ea6 DiT \u751f\u6210\u6642\uff0c\u6700\u5f8c\u968e\u6bb5\u7684 VAE \u89e3\u78bc\u5f80\u5f80\u6703\u7522\u751f\u7a81\u767c\u6027\u986f\u5b58\u5cf0\u503c\uff08VAE Spike\uff09\uff0c\u82e5\u986f\u5b58\u4e0d\u8db3\u5efa\u8b70\u958b\u555f\u5206\u584a\u89e3\u78bc\uff08Tiled VAE Decode\uff09\u4ee5\u9632\u7a0b\u5f0f\u5d29\u6f70\u3002<\/p>\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u751f\u6210\u5f0f AI \u67b6\u69cb Transformer LLM DiT \u64f4\u6563\u6a21\u578b RTX 4090 \u6548\u80fd\u76e3\u63a7 \u786c\u9ad4\u74f6\u9838\u5256\u6790 \u751f\u6210 1 \u500b Token \u7684\u4ee3\u50f9\uff1aLLM \u8207 DiT \u7b97\u529b\u74f6\u9838\u5256\u6790\u3001RTX 4090 \u5be6\u8e10\u8207\u95dc\u9375\u76e3\u63a7\u6307\u6a19\u6821\u6b63 \u5728\u7576\u524d\u751f\u6210\u5f0f AI \u67b6\u69cb\u4e2d\uff0c\u81ea\u56de\u6b78\u5927\u578b\u8a9e\u8a00\u6a21\u578b\uff08Autoregressive LLM\uff09\u8207\u64f4\u6563 Transformer\uff08DiT\uff09\u96d6\u7136\u5e95\u5c64\u7686\u57fa\u65bc\u6ce8\u610f\u529b\u6a5f\u5236\uff08Self-Attention\uff09\uff0c\u4f46\u5728\u300c\u7522\u751f 1 \u500b Token\u300d\u7684\u7269\u7406\u610f\u6db5\u3001\u786c\u9ad4\u8cc7\u6e90\u74f6\u9838\u8207\u7522\u51fa\u4ee3\u50f9\u4e0a\u5b58\u5728\u672c\u8cea\u5dee\u7570\u3002\u672c\u6587\u9664\u4e86\u6df1\u5165\u5256\u6790\u5169\u8005\u7684\u67b6\u69cb\u958b\u92b7\uff0c\u66f4\u63d0\u4f9b\u4e00\u5957\u53ef\u5728\u6d88\u8cbb\u7d1a\u65d7\u8266\u986f\u793a\u5361\uff08NVIDIA GeForce RTX 4090\uff09\u4e0a\u7a69\u5b9a\u904b\u884c\u7684\u5be6\u9ad4\u76e3\u63a7\u8173\u672c\uff0c\u4e26\u7279\u5225\u91dd\u5c0d\u696d\u754c\u6700\u5e38\u8aa4\u8b80\u7684 utilization.memory \u6838\u5fc3\u6307\u6a19\u9032\u884c\u6df1\u5ea6\u91d0\u6e05\u3002 \u76ee\u9304\u5c0e\u89bd 1. \u6982\u5ff5\u89e3\u69cb\uff1a\u96e2\u6563\u6587\u5b57 Token vs \u7a7a\u9593 Patch Token 2. \u786c\u9ad4\u67b6\u69cb\u74f6\u9838\uff1a\u8a18\u61b6\u9ad4\u983b\u5bec vs \u7b97\u529b\u5bc6\u96c6 3. \u6210\u672c\u5c0d\u6c7a\uff1a\u55ae Token \u5230\u53ef\u7528\u8f38\u51fa\u7684\u7e3d\u91cf\u5206\u6790 4. RTX 4090 \u5be6\u9ad4\u76e3\u63a7\u8173\u672c\u8207\u6307\u6a19\u5b9a\u7fa9\u6821\u6b63 5. [&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-12654","page","type-page","status-publish","hentry"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/pages\/12654","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=12654"}],"version-history":[{"count":7,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/pages\/12654\/revisions"}],"predecessor-version":[{"id":12665,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/pages\/12654\/revisions\/12665"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=12654"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}