{"id":30889,"date":"2025-05-06T12:12:40","date_gmt":"2025-05-06T04:12:40","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=30889"},"modified":"2025-05-06T12:12:40","modified_gmt":"2025-05-06T04:12:40","slug":"mimo","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/pt\/mimo\/","title":{"rendered":"MiMo\uff1a\u9ad8\u6548\u6570\u5b66\u63a8\u7406\u4e0e\u4ee3\u7801\u751f\u6210\u7684\u5c0f\u578b\u5f00\u6e90\u6a21\u578b"},"content":{"rendered":"<p>MiMo \u662f\u5c0f\u7c73\u516c\u53f8\u5f00\u53d1\u7684\u4e00\u4e2a\u5f00\u6e90\u5927\u8bed\u8a00\u6a21\u578b\u9879\u76ee\uff0c\u4e13\u6ce8\u4e8e\u6570\u5b66\u63a8\u7406\u548c\u4ee3\u7801\u751f\u6210\u3002\u6838\u5fc3\u4ea7\u54c1\u662f MiMo-7B \u7cfb\u5217\u6a21\u578b\uff0c\u5305\u542b\u57fa\u7840\u6a21\u578b (Base)\u3001\u76d1\u7763\u5fae\u8c03\u6a21\u578b (SFT)\u3001\u4ece\u57fa\u7840\u6a21\u578b\u8bad\u7ec3\u7684\u5f3a\u5316\u5b66\u4e60\u6a21\u578b (RL-Zero) \u548c\u4ece SFT \u6a21\u578b\u8bad\u7ec3\u7684\u5f3a\u5316\u5b66\u4e60\u6a21\u578b (RL)\u3002\u8fd9\u4e9b 70 \u4ebf\u53c2\u6570\u6a21\u578b\u901a\u8fc7\u4f18\u5316\u9884\u8bad\u7ec3\u6570\u636e\u3001\u591a\u91cd\u4ee4\u724c\u9884\u6d4b (MTP) \u548c\u5f3a\u5316\u5b66\u4e60\uff0c\u5c55\u73b0\u51fa\u5ab2\u7f8e\u66f4\u5927\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002MiMo-7B-RL \u5728\u6570\u5b66\u548c\u4ee3\u7801\u4efb\u52a1\u4e2d\u6027\u80fd\u7a81\u51fa\uff0c\u53ef\u5339\u654c OpenAI o1-mini\u3002\u6a21\u578b\u652f\u6301 <a href=\"https:\/\/www.kdjingpai.com\/de\/vllm\/\">vLLM<\/a> \u548c SGLang \u63a8\u7406\u5f15\u64ce\uff0c\u5e76\u5728 Hugging Face \u548c ModelScope \u63d0\u4f9b\u4e0b\u8f7d\u3002\u5c0f\u7c73\u5f00\u6e90 MiMo\uff0c\u65e8\u5728\u63a8\u52a8\u9ad8\u6548\u63a8\u7406\u6a21\u578b\u7684\u53d1\u5c55\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-30890\" title=\"MiMo\uff1a\u9ad8\u6548\u6570\u5b66\u63a8\u7406\u4e0e\u4ee3\u7801\u751f\u6210\u7684\u5c0f\u578b\u5f00\u6e90\u6a21\u578b-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/05\/547d36e8c9c2120.jpg\" alt=\"MiMo\uff1a\u9ad8\u6548\u6570\u5b66\u63a8\u7406\u4e0e\u4ee3\u7801\u751f\u6210\u7684\u5c0f\u578b\u5f00\u6e90\u6a21\u578b-1\" width=\"2366\" height=\"1207\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/05\/547d36e8c9c2120.jpg 2366w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/05\/547d36e8c9c2120-768x392.jpg 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/05\/547d36e8c9c2120-1536x784.jpg 1536w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/05\/547d36e8c9c2120-2048x1045.jpg 2048w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/05\/547d36e8c9c2120-18x9.jpg 18w\" sizes=\"auto, (max-width: 2366px) 100vw, 2366px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<ul>\n<li><strong>\u6570\u5b66\u63a8\u7406<\/strong>\uff1a\u89e3\u51b3 AIME\u3001MATH-500 \u7b49\u6570\u5b66\u7ade\u8d5b\u9898\u76ee\uff0c\u652f\u6301\u590d\u6742\u95ee\u9898\u63a8\u7406\u3002<\/li>\n<li><strong>\u4ee3\u7801\u751f\u6210<\/strong>\uff1a\u751f\u6210 Python\u3001C++ \u7b49\u9ad8\u8d28\u91cf\u4ee3\u7801\uff0c\u9002\u7528\u4e8e LiveCodeBench \u7f16\u7a0b\u4efb\u52a1\u3002<\/li>\n<li><strong>\u591a\u91cd\u4ee4\u724c\u9884\u6d4b (MTP)<\/strong>\uff1a\u9884\u6d4b\u591a\u4e2a\u4ee4\u724c\uff0c\u63a8\u7406\u63a5\u53d7\u7387\u7ea6 90%\uff0c\u63d0\u5347\u901f\u5ea6\u548c\u51c6\u786e\u6027\u3002<\/li>\n<li><strong>\u5f00\u6e90\u6a21\u578b\u652f\u6301<\/strong>\uff1a\u63d0\u4f9b MiMo-7B \u7cfb\u5217\u6a21\u578b (Base\u3001SFT\u3001RL-Zero\u3001RL)\uff0c\u4f9b\u5f00\u53d1\u8005\u81ea\u7531\u4f7f\u7528\u3002<\/li>\n<li><strong>\u9ad8\u6548\u63a8\u7406\u5f15\u64ce<\/strong>\uff1a\u652f\u6301\u5c0f\u7c73\u5b9a\u5236 vLLM \u548c SGLang\uff0c\u4f18\u5316\u63a8\u7406\u6027\u80fd\u3002<\/li>\n<li><strong>\u5f3a\u5316\u5b66\u4e60\u4f18\u5316<\/strong>\uff1a\u57fa\u4e8e 13 \u4e07\u6570\u5b66\u548c\u4ee3\u7801\u95ee\u9898\u6570\u636e\u96c6\uff0c\u63d0\u5347\u6a21\u578b\u63a8\u7406\u80fd\u529b\u3002<\/li>\n<li><strong>\u65e0\u7f1d\u56de\u6eda\u5f15\u64ce<\/strong>\uff1a\u52a0\u901f\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\uff0c\u8bad\u7ec3\u901f\u5ea6\u63d0\u5347 2.29 \u500d\uff0c\u9a8c\u8bc1\u901f\u5ea6\u63d0\u5347 1.96 \u500d\u3002<\/li>\n<li><strong>\u7075\u6d3b\u90e8\u7f72<\/strong>\uff1a\u652f\u6301 Hugging Face Transformers\u3001vLLM \u548c SGLang \u591a\u79cd\u90e8\u7f72\u65b9\u5f0f\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<h3>\u5b89\u88c5\u4e0e\u90e8\u7f72<\/h3>\n<p>MiMo-7B \u6a21\u578b\u65e0\u9700\u72ec\u7acb\u8f6f\u4ef6\u5b89\u88c5\uff0c\u4f46\u9700\u914d\u7f6e\u63a8\u7406\u73af\u5883\u3002\u4ee5\u4e0b\u4e3a\u8be6\u7ec6\u90e8\u7f72\u6b65\u9aa4\uff0c\u63a8\u8350\u4f7f\u7528 Python 3.8 \u6216\u66f4\u9ad8\u7248\u672c\u3002<\/p>\n<h4>1. \u73af\u5883\u51c6\u5907<\/h4>\n<p>\u786e\u4fdd\u7cfb\u7edf\u5b89\u88c5\u4e86 Python \u548c pip\u3002\u5efa\u8bae\u4f7f\u7528\u865a\u62df\u73af\u5883\u4ee5\u907f\u514d\u4f9d\u8d56\u51b2\u7a81\uff1a<\/p>\n<pre><code>python3 -m venv mimo_env\r\nsource mimo_env\/bin\/activate\r\n<\/code><\/pre>\n<h4>2. \u5b89\u88c5\u4f9d\u8d56<\/h4>\n<p>MiMo \u63a8\u8350\u4f7f\u7528\u5c0f\u7c73\u5b9a\u5236\u7684 vLLM \u5206\u652f\uff0c\u652f\u6301 MTP \u529f\u80fd\u3002\u5b89\u88c5\u547d\u4ee4\u5982\u4e0b\uff1a<\/p>\n<pre><code>pip install torch transformers\r\npip install \"vllm @ git+https:\/\/github.com\/XiaomiMiMo\/vllm.git@feat_mimo_mtp_stable_073\"\r\n<\/code><\/pre>\n<p>\u82e5\u4f7f\u7528 SGLang\uff0c\u6267\u884c\uff1a<\/p>\n<pre><code>python3 -m pip install \"sglang[all] @ git+https:\/\/github.com\/sgl-project\/sglang.git@main#egg=sglang&amp;subdirectory=python\"\r\n<\/code><\/pre>\n<h4>3. \u4e0b\u8f7d\u6a21\u578b<\/h4>\n<p>MiMo-7B \u6a21\u578b\u6258\u7ba1\u5728 Hugging Face \u548c ModelScope\u3002\u4ee5\u4e0b\u4ee5 MiMo-7B-RL \u4e3a\u4f8b\uff1a<\/p>\n<pre><code>from transformers import AutoModelForCausalLM, AutoTokenizer\r\nmodel_id = \"XiaomiMiMo\/MiMo-7B-RL\"\r\nmodel = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)\r\ntokenizer = AutoTokenizer.from_pretrained(model_id)\r\n<\/code><\/pre>\n<p>\u6a21\u578b\u6587\u4ef6\u7ea6 14GB\uff0c\u786e\u4fdd\u6709\u8db3\u591f\u5b58\u50a8\u7a7a\u95f4\u3002ModelScope \u4e0b\u8f7d\u65b9\u5f0f\u7c7b\u4f3c\uff0c\u66ff\u6362\u00a0<code>model_id<\/code>\u00a0\u4e3a\u5bf9\u5e94\u5730\u5740\u3002<\/p>\n<h4>4. \u542f\u52a8\u63a8\u7406\u670d\u52a1<\/h4>\n<p>\u4f7f\u7528 vLLM \u542f\u52a8\u63a8\u7406\u670d\u52a1\u5668\uff08\u63a8\u8350\uff09\uff1a<\/p>\n<pre><code>python3 -m vllm.entrypoints.api_server --model XiaomiMiMo\/MiMo-7B-RL --host 0.0.0.0 --trust-remote-code\r\n<\/code><\/pre>\n<p>\u6216\u4f7f\u7528 SGLang\uff1a<\/p>\n<pre><code>python3 -m sglang.launch_server --model-path XiaomiMiMo\/MiMo-7B-RL --host 0.0.0.0 --trust-remote-code\r\n<\/code><\/pre>\n<p>\u670d\u52a1\u5668\u542f\u52a8\u540e\uff0c\u53ef\u901a\u8fc7 API \u6216\u547d\u4ee4\u884c\u4e0e\u6a21\u578b\u4ea4\u4e92\u3002<\/p>\n<h3>\u4e3b\u8981\u529f\u80fd\u64cd\u4f5c<\/h3>\n<h4>\u6570\u5b66\u63a8\u7406<\/h4>\n<p>MiMo-7B-RL \u5728\u6570\u5b66\u63a8\u7406\u4efb\u52a1\u4e2d\u8868\u73b0\u5353\u8d8a\uff0c\u7279\u522b\u662f\u5728 AIME \u548c MATH-500 \u6570\u636e\u96c6\u4e0a\u3002\u7528\u6237\u53ef\u8f93\u5165\u6570\u5b66\u95ee\u9898\uff0c\u6a21\u578b\u751f\u6210\u89e3\u7b54\u3002\u4f8b\u5982\uff1a<\/p>\n<pre><code>from vllm import LLM, SamplingParams\r\nllm = LLM(model=\"XiaomiMiMo\/MiMo-7B-RL\", trust_remote_code=True)\r\nsampling_params = SamplingParams(temperature=0.6)\r\noutputs = llm.generate([\"Solve: 2x + 3 = 7\"], sampling_params)\r\nprint(outputs[0].outputs[0].text)\r\n<\/code><\/pre>\n<p><strong>\u64cd\u4f5c\u63d0\u793a<\/strong>\uff1a<\/p>\n<ul>\n<li>\u4f7f\u7528\u00a0<code>temperature=0.6<\/code>\u00a0\u5e73\u8861\u751f\u6210\u8d28\u91cf\u548c\u591a\u6837\u6027\u3002<\/li>\n<li>\u590d\u6742\u95ee\u9898\u53ef\u5206\u6b65\u8f93\u5165\uff0c\u786e\u4fdd\u63cf\u8ff0\u6e05\u6670\u3002<\/li>\n<li>\u652f\u6301 AIME 2024 (68.2% Pass@1)\u3001AIME 2025 (55.4% Pass@1) \u548c MATH-500 (95.8% Pass@1)\u3002<\/li>\n<\/ul>\n<h4>\u4ee3\u7801\u751f\u6210<\/h4>\n<p>MiMo-7B-RL \u53ef\u751f\u6210\u9ad8\u8d28\u91cf\u4ee3\u7801\uff0c\u652f\u6301 Python\u3001C++ \u7b49\u8bed\u8a00\uff0c\u9002\u7528\u4e8e LiveCodeBench v5 (57.8% Pass@1) \u548c v6 (49.3% Pass@1)\u3002\u793a\u4f8b\uff1a<\/p>\n<pre><code>from vllm import LLM, SamplingParams\r\nllm = LLM(model=\"XiaomiMiMo\/MiMo-7B-RL\", trust_remote_code=True)\r\nsampling_params = SamplingParams(temperature=0.6)\r\noutputs = llm.generate([\"Write a Python function to calculate factorial\"], sampling_params)\r\nprint(outputs[0].outputs[0].text)\r\n<\/code><\/pre>\n<p><strong>\u64cd\u4f5c\u63d0\u793a<\/strong>\uff1a<\/p>\n<ul>\n<li>\u63d0\u4f9b\u5177\u4f53\u4efb\u52a1\u63cf\u8ff0\uff0c\u5982\u51fd\u6570\u8f93\u5165\u8f93\u51fa\u8981\u6c42\u3002<\/li>\n<li>\u68c0\u67e5\u751f\u6210\u4ee3\u7801\u7684\u8bed\u6cd5\u5b8c\u6574\u6027\u3002<\/li>\n<li>\u9002\u5408\u7b97\u6cd5\u8bbe\u8ba1\u548c\u7f16\u7a0b\u7ade\u8d5b\u4efb\u52a1\u3002<\/li>\n<\/ul>\n<h4>\u591a\u91cd\u4ee4\u724c\u9884\u6d4b (MTP)<\/h4>\n<p>MTP \u662f MiMo \u7684\u6838\u5fc3\u7279\u8272\uff0c\u901a\u8fc7\u9884\u6d4b\u591a\u4e2a\u4ee4\u724c\u52a0\u901f\u63a8\u7406\uff0c\u63a5\u53d7\u7387\u7ea6 90%\u3002\u542f\u7528 MTP \u9700\u4f7f\u7528\u5c0f\u7c73\u5b9a\u5236 vLLM\uff1a<\/p>\n<pre><code>from vllm import LLM, SamplingParams\r\nllm = LLM(model=\"XiaomiMiMo\/MiMo-7B-RL\", trust_remote_code=True, num_speculative_tokens=1)\r\nsampling_params = SamplingParams(temperature=0.6)\r\noutputs = llm.generate([\"Write a Python script\"], sampling_params)\r\nprint(outputs[0].outputs[0].text)\r\n<\/code><\/pre>\n<p><strong>\u64cd\u4f5c\u63d0\u793a<\/strong>\uff1a<\/p>\n<ul>\n<li>\u8bbe\u7f6e\u00a0<code>num_speculative_tokens=1<\/code>\u00a0\u542f\u7528 MTP\u3002<\/li>\n<li>MTP \u5728\u9ad8\u541e\u5410\u91cf\u573a\u666f\u4e0b\u6548\u679c\u6700\u4f73\u3002<\/li>\n<li>MTP \u5c42\u5728\u9884\u8bad\u7ec3\u548c SFT \u9636\u6bb5\u8c03\u6574\uff0cRL \u9636\u6bb5\u51bb\u7ed3\u3002<\/li>\n<\/ul>\n<h4>\u65e0\u7f1d\u56de\u6eda\u5f15\u64ce<\/h4>\n<p>MiMo \u5f00\u53d1\u4e86\u65e0\u7f1d\u56de\u6eda\u5f15\u64ce\uff0c\u4f18\u5316\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u3002\u7528\u6237\u65e0\u9700\u76f4\u63a5\u64cd\u4f5c\u6b64\u529f\u80fd\uff0c\u4f46\u5176\u6548\u679c\u4f53\u73b0\u5728\u6a21\u578b\u6027\u80fd\u4e0a\uff1a<\/p>\n<ul>\n<li>\u8bad\u7ec3\u901f\u5ea6\u63d0\u5347 2.29 \u500d\uff0c\u9a8c\u8bc1\u901f\u5ea6\u63d0\u5347 1.96 \u500d\u3002<\/li>\n<li>\u96c6\u6210\u8fde\u7eed\u56de\u6eda\u3001\u5f02\u6b65\u5956\u52b1\u8ba1\u7b97\u548c\u65e9\u671f\u7ec8\u6b62\uff0c\u51cf\u5c11 GPU \u7a7a\u95f2\u65f6\u95f4\u3002<\/li>\n<\/ul>\n<h3>\u63a8\u7406\u5f15\u64ce\u9009\u62e9<\/h3>\n<ul>\n<li><strong>vLLM\uff08\u63a8\u8350\uff09<\/strong>\uff1a\u5c0f\u7c73\u5b9a\u5236 vLLM\uff08\u57fa\u4e8e vLLM 0.7.3\uff09\u652f\u6301 MTP\uff0c\u6027\u80fd\u6700\u4f18\u3002\u9002\u5408\u9ad8\u6027\u80fd\u63a8\u7406\u9700\u6c42\u3002<\/li>\n<li><strong>SGLang<\/strong>\uff1a\u652f\u6301\u4e3b\u6d41\u63a8\u7406\uff0cMTP \u652f\u6301\u5373\u5c06\u4e0a\u7ebf\u3002\u9002\u5408\u5feb\u901f\u90e8\u7f72\u3002<\/li>\n<li><strong>Hugging Face Transformers<\/strong>\uff1a\u9002\u5408\u7b80\u5355\u6d4b\u8bd5\u6216\u672c\u5730\u8c03\u8bd5\uff0c\u4f46\u4e0d\u652f\u6301 MTP\u3002<\/li>\n<\/ul>\n<h3>\u6ce8\u610f\u4e8b\u9879<\/h3>\n<ul>\n<li><strong>\u7cfb\u7edf\u63d0\u793a<\/strong>\uff1a\u63a8\u8350\u4f7f\u7528\u7a7a\u7cfb\u7edf\u63d0\u793a\u4ee5\u83b7\u5f97\u6700\u4f73\u6027\u80fd\u3002<\/li>\n<li><strong>\u786c\u4ef6\u8981\u6c42<\/strong>\uff1a\u5efa\u8bae\u4f7f\u7528\u5355 GPU\uff08\u5982 NVIDIA A100 40GB\uff09\uff0cCPU \u63a8\u7406\u9700\u81f3\u5c11 32GB \u5185\u5b58\u3002<\/li>\n<li><strong>\u8bc4\u4f30\u8bbe\u7f6e<\/strong>\uff1a\u6240\u6709\u8bc4\u4f30\u4f7f\u7528\u00a0<code>temperature=0.6<\/code>\u3002AIME \u548c LiveCodeBench \u91c7\u7528\u591a\u6b21\u8fd0\u884c\u5e73\u5747\u503c\u3002<\/li>\n<li><strong>\u793e\u533a\u652f\u6301<\/strong>\uff1a\u5982\u9047\u95ee\u9898\uff0c\u53ef\u5728 GitHub \u63d0\u4ea4 issue \u6216\u8054\u7cfb\u00a0<code>mimo@xiaomi.com<\/code>\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>\u5e94\u7528\u573a\u666f<\/h2>\n<ol>\n<li><strong>\u5b66\u672f\u7814\u7a76<\/strong><br \/>\nMiMo-7B \u6a21\u578b\u9002\u5408\u7814\u7a76\u4eba\u5458\u63a2\u7d22\u6570\u5b66\u63a8\u7406\u548c\u4ee3\u7801\u751f\u6210\u7b97\u6cd5\u3002\u5f00\u53d1\u8005\u53ef\u57fa\u4e8e\u5f00\u6e90\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u7814\u7a76\u9884\u8bad\u7ec3\u548c\u5f3a\u5316\u5b66\u4e60\u7b56\u7565\u3002<\/li>\n<li><strong>\u7f16\u7a0b\u6559\u80b2<\/strong><br \/>\n\u6559\u5e08\u53ef\u4f7f\u7528 MiMo \u751f\u6210\u7f16\u7a0b\u7ec3\u4e60\u9898\u89e3\u7b54\uff0c\u5b66\u751f\u53ef\u9a8c\u8bc1\u4ee3\u7801\u903b\u8f91\u6216\u5b66\u4e60\u7b97\u6cd5\u5b9e\u73b0\u3002<\/li>\n<li><strong>\u7ade\u8d5b\u8bad\u7ec3<\/strong><br \/>\nMiMo \u652f\u6301 AIME \u548c MATH-500 \u6570\u5b66\u7ade\u8d5b\u9898\u76ee\uff0c\u9002\u5408\u5b66\u751f\u5907\u6218\u6570\u5b66\u548c\u7f16\u7a0b\u7ade\u8d5b\u3002<\/li>\n<li><strong>AI \u5f00\u53d1<\/strong><br \/>\n\u5f00\u53d1\u8005\u53ef\u57fa\u4e8e MiMo-7B \u6784\u5efa\u5b9a\u5236\u5316\u5e94\u7528\uff0c\u5982\u81ea\u52a8\u5316\u4ee3\u7801\u5ba1\u67e5\u5de5\u5177\u6216\u6570\u5b66\u6c42\u89e3\u5668\u3002<\/li>\n<\/ol>\n<p>&nbsp;<\/p>\n<h2>QA<\/h2>\n<ol>\n<li><strong>MiMo-7B \u7cfb\u5217\u6709\u54ea\u4e9b\u6a21\u578b\uff1f<\/strong><br \/>\nMiMo-7B \u5305\u62ec\u57fa\u7840\u6a21\u578b (Base)\u3001\u76d1\u7763\u5fae\u8c03\u6a21\u578b (SFT)\u3001\u4ece\u57fa\u7840\u6a21\u578b\u8bad\u7ec3\u7684\u5f3a\u5316\u5b66\u4e60\u6a21\u578b (RL-Zero) \u548c\u4ece SFT \u6a21\u578b\u8bad\u7ec3\u7684\u5f3a\u5316\u5b66\u4e60\u6a21\u578b (RL)\u3002RL \u7248\u672c\u6027\u80fd\u6700\u4f73\u3002<\/li>\n<li><strong>\u5982\u4f55\u9009\u62e9\u63a8\u7406\u5f15\u64ce\uff1f<\/strong><br \/>\n\u63a8\u8350\u5c0f\u7c73\u5b9a\u5236 vLLM\uff08\u652f\u6301 MTP\uff0c\u6027\u80fd\u6700\u4f18\uff09\u3002SGLang \u9002\u5408\u5feb\u901f\u90e8\u7f72\uff0cHugging Face Transformers \u9002\u5408\u7b80\u5355\u6d4b\u8bd5\u3002<\/li>\n<li><strong>MTP \u5982\u4f55\u63d0\u5347\u6027\u80fd\uff1f<\/strong><br \/>\nMTP \u901a\u8fc7\u9884\u6d4b\u591a\u4e2a\u4ee4\u724c\uff0c\u63a8\u7406\u63a5\u53d7\u7387\u8fbe 90%\uff0c\u663e\u8457\u63d0\u5347\u901f\u5ea6\uff0c\u9002\u5408\u9ad8\u541e\u5410\u91cf\u573a\u666f\u3002<\/li>\n<li><strong>\u6a21\u578b\u652f\u6301\u591a\u8bed\u8a00\u5417\uff1f<\/strong><br \/>\nMiMo \u4e3b\u8981\u4f18\u5316\u6570\u5b66\u548c\u4ee3\u7801\u4efb\u52a1\uff0c\u652f\u6301\u82f1\u6587\u548c\u90e8\u5206\u4e2d\u6587\u8f93\u5165\uff0c\u672a\u660e\u786e\u652f\u6301\u5176\u4ed6\u8bed\u8a00\u3002<\/li>\n<li><strong>\u786c\u4ef6\u8981\u6c42\u662f\u4ec0\u4e48\uff1f<\/strong><br \/>\n\u5355 GPU\uff08\u5982 NVIDIA A100 40GB\uff09\u53ef\u8fd0\u884c MiMo-7B-RL\u3002CPU \u63a8\u7406\u9700\u81f3\u5c11 32GB \u5185\u5b58\uff0c\u4f46\u901f\u5ea6\u8f83\u6162\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>MiMo \u662f\u5c0f\u7c73\u516c\u53f8\u5f00\u53d1\u7684\u4e00\u4e2a\u5f00\u6e90\u5927\u8bed\u8a00\u6a21\u578b\u9879\u76ee\uff0c\u4e13\u6ce8\u4e8e\u6570\u5b66\u63a8\u7406\u548c\u4ee3\u7801\u751f\u6210\u3002\u6838\u5fc3\u4ea7\u54c1\u662f MiMo-7B \u7cfb\u5217\u6a21\u578b\uff0c\u5305\u542b\u57fa\u7840\u6a21\u578b (Base)\u3001\u76d1\u7763\u5fae\u8c03\u6a21\u578b (SFT)\u3001\u4ece\u57fa\u7840\u6a21\u578b\u8bad\u7ec3\u7684\u5f3a\u5316\u5b66\u4e60\u6a21\u578b (RL-Zero) \u548c\u4ece SFT \u6a21\u578b\u8bad\u7ec3&#8230;<\/p>\n","protected":false},"author":1,"featured_media":62325,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[230],"class_list":["post-30889","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","tag-aikaiyuanxiangmu"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/posts\/30889","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/comments?post=30889"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/posts\/30889\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/media\/62325"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/media?parent=30889"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/categories?post=30889"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/tags?post=30889"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}