{"id":33772,"date":"2025-07-26T17:43:46","date_gmt":"2025-07-26T09:43:46","guid":{"rendered":"https:\/\/www.kdjingpai.com\/?p=33772"},"modified":"2025-07-26T17:43:46","modified_gmt":"2025-07-26T09:43:46","slug":"qwen3-235b-a22b-thinking-2507","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/de\/qwen3-235b-a22b-thinking-2507\/","title":{"rendered":"Qwen3-235B-A22B-Thinking-2507\uff1a\u652f\u6301\u590d\u6742\u63a8\u7406\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b"},"content":{"rendered":"<p>Qwen3-235B-A22B-Thinking-2507 \u662f\u7531\u963f\u91cc\u5df4\u5df4\u4e91 Qwen \u56e2\u961f\u5f00\u53d1\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff0c\u53d1\u5e03\u4e8e2025\u5e747\u670825\u65e5\uff0c\u6258\u7ba1\u5728 Hugging Face \u5e73\u53f0\u4e0a\u3002\u5b83\u4e13\u6ce8\u4e8e\u590d\u6742\u63a8\u7406\u4efb\u52a1\uff0c\u652f\u6301\u9ad8\u8fbe256K\uff08262,144\uff09\u4e2a\u4ee4\u724c\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\uff0c\u9002\u5408\u5904\u7406\u903b\u8f91\u63a8\u7406\u3001\u6570\u5b66\u3001\u79d1\u5b66\u3001\u7f16\u7a0b\u548c\u5b66\u672f\u4efb\u52a1\u3002\u6a21\u578b\u91c7\u7528\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09\u67b6\u6784\uff0c\u62e5\u67092350\u4ebf\u4e2a\u53c2\u6570\uff0c\u6bcf\u6b21\u63a8\u7406\u6fc0\u6d3b220\u4ebf\u4e2a\u53c2\u6570\uff0c\u517c\u987e\u6027\u80fd\u4e0e\u6548\u7387\u3002\u5b83\u5728\u5f00\u6e90\u63a8\u7406\u6a21\u578b\u4e2d\u8868\u73b0\u7a81\u51fa\uff0c\u7279\u522b\u9002\u5408\u9700\u8981\u6df1\u5165\u601d\u8003\u548c\u957f\u4e0a\u4e0b\u6587\u5904\u7406\u7684\u5e94\u7528\u573a\u666f\u3002\u7528\u6237\u53ef\u901a\u8fc7\u591a\u79cd\u63a8\u7406\u6846\u67b6\u5982 transformers\u3001sglang \u548c <a href=\"https:\/\/www.kdjingpai.com\/en\/vllm\/\">vLLM<\/a> \u90e8\u7f72\u6a21\u578b\uff0c\u4e5f\u652f\u6301\u672c\u5730\u8fd0\u884c\u3002<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<ul>\n<li>\u652f\u6301256K\u4ee4\u724c\u7684\u8d85\u957f\u4e0a\u4e0b\u6587\u7406\u89e3\uff0c\u5904\u7406\u590d\u6742\u6587\u6863\u6216\u591a\u8f6e\u5bf9\u8bdd\u3002<\/li>\n<li>\u63d0\u4f9b\u5f3a\u5927\u7684\u903b\u8f91\u63a8\u7406\u80fd\u529b\uff0c\u9002\u7528\u4e8e\u6570\u5b66\u3001\u79d1\u5b66\u548c\u5b66\u672f\u95ee\u9898\u3002<\/li>\n<li>\u64c5\u957f\u7f16\u7a0b\u4efb\u52a1\uff0c\u652f\u6301\u4ee3\u7801\u751f\u6210\u548c\u8c03\u8bd5\u3002<\/li>\n<li>\u96c6\u6210\u5de5\u5177\u8c03\u7528\u529f\u80fd\uff0c\u901a\u8fc7 Qwen-Agent \u7b80\u5316\u5916\u90e8\u5de5\u5177\u4ea4\u4e92\u3002<\/li>\n<li>\u652f\u6301100\u591a\u79cd\u8bed\u8a00\uff0c\u9002\u5408\u591a\u8bed\u8a00\u6307\u4ee4\u9075\u5faa\u548c\u7ffb\u8bd1\u3002<\/li>\n<li>\u63d0\u4f9b FP8 \u91cf\u5316\u7248\u672c\uff0c\u964d\u4f4e\u786c\u4ef6\u9700\u6c42\uff0c\u4f18\u5316\u63a8\u7406\u6027\u80fd\u3002<\/li>\n<li>\u517c\u5bb9\u591a\u79cd\u63a8\u7406\u6846\u67b6\uff0c\u5982 transformers\u3001sglang\u3001vLLM \u548c llama.cpp\u3002<\/li>\n<\/ul>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<h3>\u5b89\u88c5\u4e0e\u90e8\u7f72<\/h3>\n<p>\u8981\u4f7f\u7528 Qwen3-235B-A22B-Thinking-2507\uff0c\u9700\u51c6\u5907\u9ad8\u6027\u80fd\u8ba1\u7b97\u73af\u5883\uff0c\u56e0\u5176\u6a21\u578b\u6587\u4ef6\u8f83\u5927\uff08BF16\u7248\u672c\u7ea6437.91GB\uff0cFP8\u7248\u672c\u7ea6220.20GB\uff09\u3002\u4ee5\u4e0b\u662f\u8be6\u7ec6\u5b89\u88c5\u6b65\u9aa4\uff1a<\/p>\n<ol>\n<li><strong>\u73af\u5883\u51c6\u5907<\/strong>\uff1a\n<ul>\n<li>\u786e\u4fdd\u786c\u4ef6\u6ee1\u8db3\u8981\u6c42\uff1aBF16\u7248\u672c\u5efa\u8bae88GB\u663e\u5b58\uff0cFP8\u7248\u672c\u7ea630GB\u663e\u5b58\u3002<\/li>\n<li>\u5b89\u88c5 Python 3.8+ \u548c PyTorch\uff0c\u63a8\u8350\u4f7f\u7528 CUDA \u652f\u6301\u7684 GPU \u73af\u5883\u3002<\/li>\n<li>\u5b89\u88c5 Hugging Face transformers \u5e93\uff0c\u7248\u672c\u9700\u22654.51.0\uff0c\u907f\u514d\u517c\u5bb9\u6027\u95ee\u9898\uff1a\n<pre><code>pip install transformers&gt;=4.51.0\r\n<\/code><\/pre>\n<\/li>\n<li>\u53ef\u9009\u5b89\u88c5 sglang\uff08\u22650.4.6.post1\uff09\u6216 vLLM\uff08\u22650.8.5\uff09\u4ee5\u652f\u6301\u9ad8\u6548\u63a8\u7406\uff1a\n<pre><code>pip install sglang&gt;=0.4.6.post1 vllm&gt;=0.8.5\r\n<\/code><\/pre>\n<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u4e0b\u8f7d\u6a21\u578b<\/strong>\uff1a\n<ul>\n<li>\u4ece Hugging Face \u4ed3\u5e93\u4e0b\u8f7d\u6a21\u578b\uff1a\n<pre><code>huggingface-cli download Qwen\/Qwen3-235B-A22B-Thinking-2507\r\n<\/code><\/pre>\n<\/li>\n<li>\u82e5\u9700 FP8 \u7248\u672c\uff0c\u4e0b\u8f7d Qwen3-235B-A22B-Thinking-2507-FP8\uff1a\n<pre><code>huggingface-cli download Qwen\/Qwen3-235B-A22B-Thinking-2507-FP8\r\n<\/code><\/pre>\n<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u672c\u5730\u8fd0\u884c<\/strong>\uff1a\n<ul>\n<li>\u4f7f\u7528 transformers \u52a0\u8f7d\u6a21\u578b\uff1a\n<pre><code>from transformers import AutoModelForCausalLM, AutoTokenizer\r\nmodel_name = \"Qwen\/Qwen3-235B-A22B-Thinking-2507\"\r\ntokenizer = AutoTokenizer.from_pretrained(model_name)\r\nmodel = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=\"auto\", device_map=\"auto\")\r\n<\/code><\/pre>\n<\/li>\n<li>\u4e3a\u907f\u514d\u5185\u5b58\u4e0d\u8db3\uff0c\u53ef\u51cf\u5c0f\u4e0a\u4e0b\u6587\u957f\u5ea6\uff08\u598232768\u4ee4\u724c\uff09\uff1a\n<pre><code>python -m sglang.launch_server --model-path Qwen\/Qwen3-235B-A22B-Thinking-2507 --tp 8 --context-length 32768 --reasoning-parser <a href=\"https:\/\/www.kdjingpai.com\/en\/deepseek-r1nenglixiang\/\">deepseek-r1<\/a>\r\n<\/code><\/pre>\n<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u5de5\u5177\u8c03\u7528\u914d\u7f6e<\/strong>\uff1a\n<ul>\n<li>\u4f7f\u7528 Qwen-Agent \u7b80\u5316\u5de5\u5177\u8c03\u7528\uff1a\n<pre><code>from qwen_agent.agents import Assistant\r\nllm_cfg = {\r\n'model': 'qwen3-235b-a22b-thinking-2507',\r\n'model_type': 'qwen_dashscope'\r\n}\r\ntools = [{'mcpServers': {'time': {'command': 'uvx', 'args': ['mcp-server-time', '--local-timezone=Asia\/Shanghai']}}}]\r\nbot = Assistant(llm=llm_cfg, function_list=tools)\r\nmessages = [{'role': 'user', 'content': '\u83b7\u53d6\u5f53\u524d\u65f6\u95f4'}]\r\nfor responses in bot.run(messages=messages):\r\nprint(responses)\r\n<\/code><\/pre>\n<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<h3>\u4e3b\u8981\u529f\u80fd\u64cd\u4f5c<\/h3>\n<ul>\n<li><strong>\u590d\u6742\u63a8\u7406<\/strong>\uff1a\u6a21\u578b\u9ed8\u8ba4\u542f\u7528\u601d\u8003\u6a21\u5f0f\uff0c\u8f93\u51fa\u5305\u542b\u00a0<code>&lt;think&gt;<\/code>\u00a0\u6807\u7b7e\uff0c\u9002\u5408\u89e3\u51b3\u6570\u5b66\u6216\u903b\u8f91\u95ee\u9898\u3002\u4f8b\u5982\uff0c\u8f93\u5165\u201c\u8bc1\u660e\u8d39\u9a6c\u5c0f\u5b9a\u7406\u201d\uff0c\u6a21\u578b\u4f1a\u751f\u6210\u9010\u6b65\u63a8\u7406\u8fc7\u7a0b\u3002<\/li>\n<li><strong>\u957f\u4e0a\u4e0b\u6587\u5904\u7406<\/strong>\uff1a\u652f\u6301256K\u4ee4\u724c\uff0c\u9002\u5408\u5206\u6790\u957f\u6587\u6863\u3002\u8f93\u5165\u957f\u7bc7\u6587\u672c\u540e\uff0c\u6a21\u578b\u53ef\u63d0\u53d6\u5173\u952e\u4fe1\u606f\u6216\u56de\u7b54\u76f8\u5173\u95ee\u9898\u3002<\/li>\n<li><strong>\u7f16\u7a0b\u652f\u6301<\/strong>\uff1a\u8f93\u5165\u4ee3\u7801\u7247\u6bb5\u6216\u95ee\u9898\uff0c\u5982\u201c\u7f16\u5199 Python \u6392\u5e8f\u7b97\u6cd5\u201d\uff0c\u6a21\u578b\u751f\u6210\u5b8c\u6574\u4ee3\u7801\u5e76\u89e3\u91ca\u903b\u8f91\u3002<\/li>\n<li><strong>\u5de5\u5177\u8c03\u7528<\/strong>\uff1a\u901a\u8fc7 Qwen-Agent\uff0c\u6a21\u578b\u53ef\u8c03\u7528\u5916\u90e8\u5de5\u5177\uff0c\u5982\u83b7\u53d6\u65f6\u95f4\u6216\u6267\u884c\u7f51\u7edc\u8bf7\u6c42\uff0c\u7b80\u5316\u590d\u6742\u4efb\u52a1\u3002<\/li>\n<\/ul>\n<h3>\u6ce8\u610f\u4e8b\u9879<\/h3>\n<ul>\n<li>\u63a8\u7406\u6a21\u5f0f\u4e0b\uff0c\u5efa\u8bae\u4e0a\u4e0b\u6587\u957f\u5ea6\u2265131072\u4ee5\u786e\u4fdd\u6027\u80fd\u3002<\/li>\n<li>\u907f\u514d\u4f7f\u7528\u8d2a\u5a6a\u89e3\u7801\uff0c\u53ef\u80fd\u5bfc\u81f4\u91cd\u590d\u8f93\u51fa\u3002<\/li>\n<li>\u5bf9\u4e8e\u672c\u5730\u8fd0\u884c\uff0c\u63a8\u8350\u4f7f\u7528 <a href=\"https:\/\/www.kdjingpai.com\/en\/ollama\/\">Ollama<\/a> \u6216 LMStudio\uff0c\u4f46\u9700\u8c03\u6574\u4e0a\u4e0b\u6587\u957f\u5ea6\u4ee5\u907f\u514d\u5faa\u73af\u95ee\u9898\u3002<\/li>\n<\/ul>\n<h2>\u5e94\u7528\u573a\u666f<\/h2>\n<ol>\n<li><strong>\u5b66\u672f\u7814\u7a76<\/strong><br \/>\n\u7814\u7a76\u4eba\u5458\u53ef\u4f7f\u7528\u6a21\u578b\u5206\u6790\u957f\u7bc7\u8bba\u6587\uff0c\u63d0\u53d6\u5173\u952e\u8bba\u70b9\u6216\u9a8c\u8bc1\u6570\u5b66\u516c\u5f0f\u3002\u5176256K\u4e0a\u4e0b\u6587\u957f\u5ea6\u652f\u6301\u5904\u7406\u6574\u7bc7\u6587\u6863\uff0c\u9002\u5408\u6587\u732e\u7efc\u8ff0\u6216\u8de8\u7ae0\u8282\u5206\u6790\u3002<\/li>\n<li><strong>\u7f16\u7a0b\u5f00\u53d1<\/strong><br \/>\n\u5f00\u53d1\u8005\u53ef\u5229\u7528\u6a21\u578b\u751f\u6210\u4ee3\u7801\u3001\u8c03\u8bd5\u7a0b\u5e8f\u6216\u4f18\u5316\u7b97\u6cd5\u3002\u4f8b\u5982\uff0c\u8f93\u5165\u590d\u6742\u7b97\u6cd5\u9700\u6c42\uff0c\u6a21\u578b\u4f1a\u63d0\u4f9b\u4ee3\u7801\u5e76\u89e3\u91ca\u5b9e\u73b0\u6b65\u9aa4\u3002<\/li>\n<li><strong>\u591a\u8bed\u8a00\u7ffb\u8bd1<\/strong><br \/>\n\u4f01\u4e1a\u53ef\u4f7f\u7528\u6a21\u578b\u8fdb\u884c\u591a\u8bed\u8a00\u6587\u6863\u7ffb\u8bd1\u6216\u6307\u4ee4\u5904\u7406\uff0c\u652f\u6301100\u591a\u79cd\u8bed\u8a00\uff0c\u9002\u5408\u8de8\u56fd\u6c9f\u901a\u6216\u672c\u5730\u5316\u4efb\u52a1\u3002<\/li>\n<li><strong>\u6559\u80b2\u652f\u6301<\/strong><br \/>\n\u5b66\u751f\u548c\u6559\u5e08\u53ef\u5229\u7528\u6a21\u578b\u89e3\u7b54\u6570\u5b66\u3001\u79d1\u5b66\u95ee\u9898\u6216\u751f\u6210\u6559\u5b66\u6750\u6599\u3002\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u5e2e\u52a9\u89e3\u91ca\u590d\u6742\u6982\u5ff5\u3002<\/li>\n<\/ol>\n<h2>QA<\/h2>\n<ol>\n<li><strong>\u6a21\u578b\u652f\u6301\u54ea\u4e9b\u63a8\u7406\u6846\u67b6\uff1f<\/strong><br \/>\n\u652f\u6301 transformers\u3001sglang\u3001vLLM\u3001Ollama\u3001LMStudio \u548c llama.cpp\u3002\u63a8\u8350\u4f7f\u7528\u6700\u65b0\u7248\u672c\u4ee5\u786e\u4fdd\u517c\u5bb9\u6027\u3002<\/li>\n<li><strong>\u5982\u4f55\u5904\u7406\u5185\u5b58\u4e0d\u8db3\u95ee\u9898\uff1f<\/strong><br \/>\n\u964d\u4f4e\u4e0a\u4e0b\u6587\u957f\u5ea6\u81f332768\uff0c\u6216\u4f7f\u7528 FP8 \u7248\u672c\u4ee5\u51cf\u5c11\u663e\u5b58\u9700\u6c42\u3002\u4e5f\u53ef\u901a\u8fc7 tensor-parallel-size \u53c2\u6570\u5206\u914d\u591a GPU \u8d44\u6e90\u3002<\/li>\n<li><strong>\u5982\u4f55\u542f\u7528\u5de5\u5177\u8c03\u7528\u529f\u80fd\uff1f<\/strong><br \/>\n\u4f7f\u7528 Qwen-Agent \u914d\u7f6e\u5de5\u5177\uff0c\u5b9a\u4e49 <a href=\"https:\/\/www.kdjingpai.com\/en\/model-context-protocol\/\">MCP<\/a> \u6587\u4ef6\u6216\u5185\u7f6e\u5de5\u5177\uff0c\u6a21\u578b\u5373\u53ef\u81ea\u52a8\u8c03\u7528\u5916\u90e8\u529f\u80fd\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>Qwen3-235B-A22B-Thinking-2507 \u662f\u7531\u963f\u91cc\u5df4\u5df4\u4e91 Qwen \u56e2\u961f\u5f00\u53d1\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff0c\u53d1\u5e03\u4e8e2025\u5e747\u670825\u65e5\uff0c\u6258\u7ba1\u5728 Hugging Face \u5e73\u53f0\u4e0a\u3002\u5b83\u4e13\u6ce8\u4e8e\u590d\u6742\u63a8\u7406\u4efb\u52a1\uff0c\u652f\u6301\u9ad8\u8fbe256K\uff08262,144\uff09\u4e2a\u4ee4&#8230;<\/p>\n","protected":false},"author":1,"featured_media":31260,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20,392,397],"tags":[230],"class_list":["post-33772","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","category-models","category-text-model","tag-aikaiyuanxiangmu"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/33772","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/comments?post=33772"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/33772\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/media\/31260"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/media?parent=33772"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/categories?post=33772"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/tags?post=33772"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}