{"id":53816,"date":"2025-02-16T16:55:48","date_gmt":"2025-02-16T08:55:48","guid":{"rendered":"https:\/\/fwq.ai\/blog\/53816\/"},"modified":"2025-02-16T16:55:48","modified_gmt":"2025-02-16T08:55:48","slug":"%e7%94%a8unsloth%e8%ae%ad%e7%bb%83%e8%87%aa%e5%b7%b1%e7%9a%84r1%e6%8e%a8%e7%90%86%e6%a8%a1%e5%9e%8b","status":"publish","type":"post","link":"https:\/\/fwq.ai\/blog\/53816\/","title":{"rendered":"\u7528Unsloth\u8bad\u7ec3\u81ea\u5df1\u7684R1\u63a8\u7406\u6a21\u578b"},"content":{"rendered":"<p>\u4eca\u5929\uff0c\u6211\u4eec\u5f88\u9ad8\u5174\u5728 Unsloth \u4e2d\u5f15\u5165\u63a8\u7406\u529f\u80fd\uff01DeepSeek \u7684 R1 \u7814\u7a76\u63ed\u793a\u4e86\u4e00\u4e2a\u201c\u987f\u609f\u65f6\u523b\u201d\uff0c\u5176\u4e2d R1-Zero \u901a\u8fc7\u4f7f\u7528\u7fa4\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316 (GRPO) \u81ea\u4e3b\u5b66\u4e60\u5206\u914d\u66f4\u591a\u601d\u8003\u65f6\u95f4\u800c\u65e0\u9700\u4eba\u5de5\u53cd\u9988\u3002<\/p>\n<p>\u6211\u4eec\u589e\u5f3a\u4e86\u6574\u4e2a GRPO \u6d41\u7a0b\uff0c\u4f7f\u5176\u4f7f\u7528\u7684 VRAM \u6bd4 Hugging Face + FA2 \u5c11 80%\u3002\u8fd9\u6837\u4f60\u5c31\u53ef\u4ee5\u4f7f\u7528 Qwen2.5 (1.5B) \u5728\u4ec5 7GB \u7684 VRAM \u4e0a\u91cd\u73b0 R1-Zero \u7684\u201c\u987f\u609f\u65f6\u523b\u201d\u3002<\/p>\n<p>\u8bd5\u7528\u6211\u4eec\u7684\u514d\u8d39 G\u200b\u200bRPO \u7b14\u8bb0\u672c\uff1a \u3002\u5bf9\u4e8e\u5177\u6709\u5176\u4ed6\u6a21\u578b\uff08\u5982 Phi-4\uff09\u7684 GRPO \u7b14\u8bb0\u672c\uff0c\u8bf7\u8bbf\u95ee<\/p>\n<h2>1\u3001\u4e3b\u8981\u7ec6\u8282<\/h2>\n<ul>\n<li>\u501f\u52a9 15GB VRAM\uff0cUnsloth \u53ef\u8ba9\u4f60\u5c06\u4efb\u4f55\u9ad8\u8fbe 15B \u53c2\u6570\u7684\u6a21\u578b\uff08\u5982 Llama 3.1 (8B)\u3001Phi-4 (14B)\u3001Mistral (7B) \u6216 Qwen2.5 (7B)\uff09\u8f6c\u6362\u4e3a\u63a8\u7406\u6a21\u578b<\/li>\n<li>\u6700\u4f4e\u8981\u6c42\uff1a\u4ec5 7GB VRAM \u5c31\u8db3\u4ee5\u5728\u672c\u5730\u8bad\u7ec3\u4f60\u81ea\u5df1\u7684\u63a8\u7406\u6a21\u578b\u3002<\/li>\n<li>\u7684\u4f18\u79c0\u56e2\u961f\u8bc1\u660e\uff0c\u4f60\u53ef\u4ee5\u4f7f\u7528 Qwen2.5 (1.5B) \u5b9e\u73b0\u81ea\u5df1\u7684\u201c\u554a\u54c8\u201d\u65f6\u523b &#8211; \u4f46\u5b83\u9700\u8981 2xA100 GPU\uff08160GB VRAM\uff09\u3002\u73b0\u5728\uff0c\u4f7f\u7528 Unsloth\uff0c\u4f60\u53ea\u9700\u4f7f\u7528\u5355\u4e2a 7GB VRAM GPU \u5373\u53ef\u5b9e\u73b0\u76f8\u540c\u7684\u201c\u987f\u609f\u201d\u65f6\u523b<\/li>\n<li>\u4ee5\u524d\uff0cGRPO \u4ec5\u652f\u6301\u5b8c\u5168\u5fae\u8c03\uff0c\u4f46\u6211\u4eec\u5df2\u4f7f\u5176\u4e0e QLoRA \u548c LoRA \u914d\u5408\u4f7f\u7528<\/li>\n<li>\u8bf7\u6ce8\u610f\uff0c\u8fd9\u4e0d\u662f\u5bf9 DeepSeek \u7684 R1 \u84b8\u998f\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u4e5f\u4e0d\u662f\u4f7f\u7528 Unsloth \u5df2\u7ecf\u652f\u6301\u7684 R1 \u84b8\u998f\u6570\u636e\u8fdb\u884c\u8c03\u6574\u3002\u8fd9\u662f\u4f7f\u7528 GRPO \u5c06\u6807\u51c6\u6a21\u578b\u8f6c\u6362\u4e3a\u6210\u719f\u7684\u63a8\u7406\u6a21\u578b\u3002<\/li>\n<li>GRPO \u7684\u7528\u4f8b\u5305\u62ec\uff1a\u5982\u679c\u4f60\u60f3\u5236\u4f5c\u5e26\u6709\u5956\u52b1\u7684\u5b9a\u5236\u6a21\u578b\uff08\u4f8b\u5982\u6cd5\u5f8b\u3001\u533b\u5b66\u7b49\uff09\uff0c\u90a3\u4e48 GRPO \u53ef\u4ee5\u63d0\u4f9b\u5e2e\u52a9\u3002<\/li>\n<\/ul>\n<p>\u5982\u679c\u4f60\u6709\u8f93\u5165\u548c\u8f93\u51fa\u6570\u636e\uff08\u5982\u95ee\u9898\u548c\u7b54\u6848\uff09\uff0c\u4f46\u6ca1\u6709\u601d\u8def\u6216\u63a8\u7406\u8fc7\u7a0b\uff0cGRPO \u53ef\u4ee5\u795e\u5947\u5730\u4e3a\u4f60\u521b\u5efa\u63a8\u7406\u8fc7\u7a0b\uff01+ \u66f4\u591a<\/p>\n<h2>2\u3001GRPO +\u201c\u987f\u609f\u201d\u65f6\u523b<\/h2>\n<p>DeepSeek \u7684\u7814\u7a76\u4eba\u5458\u5728\u4f7f\u7528\u7eaf\u5f3a\u5316\u5b66\u4e60 (RL) \u8bad\u7ec3 R1-Zero \u65f6\u89c2\u5bdf\u5230\u4e86\u201c\u987f\u609f\u201d\u65f6\u523b\u3002\u8be5\u6a21\u578b\u5b66\u4f1a\u4e86\u901a\u8fc7\u91cd\u65b0\u8bc4\u4f30\u5176\u521d\u59cb\u65b9\u6cd5\u6765\u5ef6\u957f\u5176\u601d\u8003\u65f6\u95f4\uff0c\u800c\u65e0\u9700\u4efb\u4f55\u4eba\u5de5\u6307\u5bfc\u6216\u9884\u5b9a\u4e49\u6307\u4ee4\u3002<\/p>\n<p>\u5728\u4e00\u4e2a\u6d4b\u8bd5\u793a\u4f8b\u4e2d\uff0c\u5c3d\u7ba1\u6211\u4eec\u4ec5\u4f7f\u7528 GRPO \u8bad\u7ec3\u4e86 100 \u6b65\u7684 Phi-4\uff0c\u4f46\u7ed3\u679c\u5df2\u7ecf\u5f88\u660e\u663e\u4e86\u3002\u6ca1\u6709 GRPO \u7684\u6a21\u578b\u6ca1\u6709\u601d\u8003\u6807\u8bb0\uff0c\u800c\u4f7f\u7528 GRPO \u8bad\u7ec3\u7684\u6a21\u578b\u6709\u601d\u8003\u6807\u8bb0\u5e76\u4e14\u4e5f\u6709\u6b63\u786e\u7b54\u6848\u3002<\/p>\n<p>\u8fd9\u79cd\u9b54\u529b\u53ef\u4ee5\u901a\u8fc7 GRPO \u91cd\u73b0\uff0cGRPO \u662f\u4e00\u79cd RL \u7b97\u6cd5\uff0c\u53ef\u4ee5\u6709\u6548\u5730\u4f18\u5316\u54cd\u5e94\u800c\u65e0\u9700\u4ef7\u503c\u51fd\u6570\uff0c\u800c\u4e0d\u50cf\u4f9d\u8d56\u4ef7\u503c\u51fd\u6570\u7684\u8fd1\u7aef\u7b56\u7565\u4f18\u5316 (PPO)\u3002\u5728\u6211\u4eec\u7684\u7b14\u8bb0\u672c\u4e2d\uff0c\u6211\u4eec\u4f7f\u7528 GRPO \u8bad\u7ec3\u6a21\u578b\uff0c\u65e8\u5728\u8ba9\u5b83\u81ea\u4e3b\u5f00\u53d1\u81ea\u5df1\u7684\u81ea\u6211\u9a8c\u8bc1\u548c\u641c\u7d22\u80fd\u529b &#8211; \u521b\u9020\u4e00\u4e2a\u8ff7\u4f60\u201c\u987f\u609f\u65f6\u523b\u201d\u3002<\/p>\n<p>\u5de5\u4f5c\u539f\u7406\uff1a<\/p>\n<ul>\n<li>\u6a21\u578b\u751f\u6210\u54cd\u5e94\u7ec4\u3002<\/li>\n<li>\u6bcf\u4e2a\u54cd\u5e94\u90fd\u6839\u636e\u6b63\u786e\u6027\u6216\u7531\u67d0\u4e9b\u8bbe\u7f6e\u5956\u52b1\u51fd\u6570\uff08\u800c\u4e0d\u662f LLM \u5956\u52b1\u6a21\u578b\uff09\u521b\u5efa\u7684\u53e6\u4e00\u4e2a\u6307\u6807\u8fdb\u884c\u8bc4\u5206\u3002<\/li>\n<li>\u8ba1\u7b97\u7ec4\u7684\u5e73\u5747\u5206\u6570\u3002<\/li>\n<li>\u6bcf\u4e2a\u54cd\u5e94\u7684\u5206\u6570\u4e0e\u7ec4\u5e73\u5747\u503c\u8fdb\u884c\u6bd4\u8f83\u3002<\/li>\n<li>\u6a21\u578b\u5f97\u5230\u5f3a\u5316\uff0c\u4ee5\u652f\u6301\u5f97\u5206\u66f4\u9ad8\u7684\u54cd\u5e94\u3002<\/li>\n<\/ul>\n<p>\u4f8b\u5982\uff0c\u5047\u8bbe\u6211\u4eec\u60f3\u8981\u4e00\u4e2a\u6a21\u578b\u6765\u89e3\u51b3\uff1a<\/p>\n<ul>\n<li>1+1 \u7b49\u4e8e\u591a\u5c11\uff1f&gt;&gt; \u601d\u7ef4\u94fe\/\u8ba1\u7b97 &gt;&gt; \u7b54\u6848\u662f 2\u3002<\/li>\n<li>2+2 \u7b49\u4e8e\u591a\u5c11\uff1f&gt;&gt; \u601d\u7ef4\u94fe\/\u8ba1\u7b97 &gt;&gt; \u7b54\u6848\u662f 4\u3002<\/li>\n<\/ul>\n<p>\u6700\u521d\uff0c\u4eba\u4eec\u5fc5\u987b\u6536\u96c6\u5927\u91cf\u6570\u636e\u6765\u586b\u5145\u8ba1\u7b97\/\u601d\u8def\u8fc7\u7a0b\u3002\u4f46 GRPO\uff08DeepSeek \u4f7f\u7528\u7684\u7b97\u6cd5\uff09\u6216\u5176\u4ed6 RL \u7b97\u6cd5\u53ef\u4ee5\u5f15\u5bfc\u6a21\u578b\u81ea\u52a8\u5c55\u793a\u63a8\u7406\u80fd\u529b\u5e76\u521b\u5efa\u63a8\u7406\u8f68\u8ff9\u3002\u76f8\u53cd\uff0c\u6211\u4eec\u9700\u8981\u521b\u5efa\u826f\u597d\u7684\u5956\u52b1\u51fd\u6570\u6216\u9a8c\u8bc1\u5668\u3002\u4f8b\u5982\uff0c\u5982\u679c\u5b83\u5f97\u5230\u4e86\u6b63\u786e\u7684\u7b54\u6848\uff0c\u5c31\u7ed9\u5b83 1 \u5206\u3002\u5982\u679c\u67d0\u4e9b\u5355\u8bcd\u62fc\u5199\u9519\u8bef\uff0c\u5219\u51cf 0.1\u3002\u7b49\u7b49\uff01\u6211\u4eec\u53ef\u4ee5\u63d0\u4f9b\u8bb8\u591a\u51fd\u6570\u6765\u5956\u52b1\u8fd9\u4e2a\u8fc7\u7a0b\u3002<\/p>\n<h2>3\u3001Unsloth \u4e2d\u7684 GRPO<\/h2>\n<p>\u5982\u679c\u5728\u672c\u5730\u5c06 GRPO \u4e0e Unsloth \u4e00\u8d77\u4f7f\u7528\uff0c\u8bf7 <code>pip install diffusers<\/code>\uff0c\u56e0\u4e3a\u5b83\u662f\u4e00\u4e2a\u4f9d\u8d56\u9879\u3002<\/p>\n<p>\u7b49\u5f85\u81f3\u5c11 300 \u6b65\uff0c\u5956\u52b1\u624d\u4f1a\u771f\u6b63\u589e\u52a0\uff0c\u8bf7\u4f7f\u7528\u6700\u65b0\u7248\u672c\u7684 vLLM\u3002\u8bf7\u8bb0\u4f4f\uff0c\u6211\u4eec\u5728 Colab \u4e0a\u7684\u793a\u4f8b\u53ea\u8bad\u7ec3\u4e86\u4e00\u4e2a\u5c0f\u65f6\uff0c\u56e0\u6b64\u7ed3\u679c\u4f4e\u4e8e\u6807\u51c6\u3002\u4e3a\u4e86\u83b7\u5f97\u826f\u597d\u7684\u7ed3\u679c\uff0c\u4f60\u9700\u8981\u8bad\u7ec3\u81f3\u5c11 12 \u4e2a\u5c0f\u65f6\uff08\u8fd9\u5c31\u662f GRPO \u7684\u5de5\u4f5c\u539f\u7406\uff09\uff0c\u4f46\u8bf7\u8bb0\u4f4f\u8fd9\u4e0d\u662f\u5f3a\u5236\u6027\u7684\uff0c\u56e0\u4e3a\u4f60\u53ef\u4ee5\u968f\u65f6\u505c\u6b62\u3002<\/p>\n<p>\u5c06 GRPO \u5e94\u7528\u4e8e\u53c2\u6570\u81f3\u5c11\u4e3a 1.5B \u7684\u6a21\u578b\uff0c\u4ee5\u6b63\u786e\u751f\u6210\u601d\u8003\u4ee4\u724c\uff0c\u56e0\u4e3a\u8f83\u5c0f\u7684\u6a21\u578b\u53ef\u80fd\u65e0\u6cd5\u505a\u5230\u8fd9\u4e00\u70b9\u3002\u5982\u679c\u4f60\u4f7f\u7528\u7684\u662f\u57fa\u7840\u6a21\u578b\uff0c\u8bf7\u786e\u4fdd\u6709\u4e00\u4e2a\u804a\u5929\u6a21\u677f\u3002GRPO \u7684\u8bad\u7ec3\u635f\u5931\u8ddf\u8e2a\u73b0\u5728\u76f4\u63a5\u5185\u7f6e\u5728 Unsloth \u4e2d\uff0c\u65e0\u9700\u4f7f\u7528 wandb \u7b49\u5916\u90e8\u5de5\u5177\u3002<\/p>\n<p>\u9664\u4e86\u6dfb\u52a0 GRPO \u652f\u6301\u5916\uff0c\u6211\u4eec\u968f\u540e\u8fd8\u652f\u6301\u5728\u7ebf DPO\u3001PPO \u548c RLOO\uff01\u66f4\u591a\u8be6\u7ec6\u4fe1\u606f\u53ef\u4ee5\u5728\u548c\u4e2d\u770b\u5230\uff0c\u5176\u4e2d\u5305\u62ec\u6709\u5173\u4ed6\u5982\u4f55\u8ba9\u5728\u7ebf DPO \u5de5\u4f5c\u7684 Github fork\u3002 \u4e2d\u4e5f\u53ef\u4ee5\u770b\u5230 Google Colab \u4e0a GRPO \u66f4\u6539\u7684\u521d\u7a3f\uff01\u4ed6\u4eec\u7684\u8d21\u732e\u8ba9\u6211\u4eec\u4e5f\u80fd\u591f\u652f\u6301\u5176\u4ed6\u57fa\u4e8e\u751f\u6210\u7684 RL \u65b9\u6cd5\u3002\u8bf7\u53c2\u89c1\u4e0b\u9762\u7684\u56fe\u8868\u6bd4\u8f83 Unsloth \u7684\u5728\u7ebf DPO VRAM \u6d88\u8017\u4e0e\u6807\u51c6 Hugging Face + FA2\u3002<\/p>\n<h2>4\u3001Unsloth x vLLM<\/h2>\n<p>20 \u500d\u541e\u5410\u91cf\uff0c50% VRAM \u8282\u7701\u3002<\/p>\n<p>\u4f60\u73b0\u5728\u53ef\u4ee5\u5728\u5fae\u8c03\u6808\u4e2d\u76f4\u63a5\u4f7f\u7528 vLLM\uff0c\u8fd9\u5141\u8bb8\u66f4\u5927\u7684\u541e\u5410\u91cf\uff0c\u5e76\u5141\u8bb8\u4f60\u540c\u65f6\u5bf9\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u548c\u63a8\u7406\uff01\u5728 1x A100 40GB \u4e0a\uff0c\u4f7f\u7528 Unsloth \u7684 Llama 3.2 3B Instruct \u7684\u52a8\u6001 4 \u4f4d\u91cf\u5316\uff0c\u9884\u8ba1\u6bcf\u79d2 4000 \u4e2a\u4ee4\u724c\u5de6\u53f3\u3002\u5728 16GB Tesla T4\uff08\u514d\u8d39 Colab GPU\uff09\u4e0a\uff0c\u4f60\u53ef\u4ee5\u83b7\u5f97 300 \u4e2a\u4ee4\u724c\/\u79d2\u3002<\/p>\n<p>\u6211\u4eec\u8fd8\u795e\u5947\u5730\u6d88\u9664\u4e86\u540c\u65f6\u52a0\u8f7d vLLM \u548c Unsloth \u65f6\u7684\u53cc\u500d\u5185\u5b58\u4f7f\u7528\u91cf\uff0c\u4ece\u800c\u4e3a Llama 3.1 8B \u8282\u7701\u4e86 5GB \u5de6\u53f3\uff0c\u4e3a Llama 3.2 3B \u8282\u7701\u4e86 3GB\uff08\u611f\u8c22 \u7684\u542f\u53d1\uff09\u3002 Unsloth \u6700\u521d\u53ef\u4ee5\u5728 1x 48GB GPU \u4e2d\u5fae\u8c03 Llama 3.3 70B \u6307\u4ee4\uff0c\u5176\u4e2d Llama 3.3 70B \u6743\u91cd\u5360\u7528 40GB VRAM\u3002\u5982\u679c\u6211\u4eec\u4e0d\u6d88\u9664\u53cc\u500d\u5185\u5b58\u4f7f\u7528\u91cf\uff0c\u90a3\u4e48\u5728\u540c\u65f6\u52a0\u8f7d Unsloth \u548c vLLM \u65f6\uff0c\u6211\u4eec\u5c06\u9700\u8981 &gt;= 80GB \u7684 VRAM\u3002<\/p>\n<p>\u4f46\u4f7f\u7528 Unsloth\uff0c\u4f60\u4ecd\u7136\u53ef\u4ee5\u5728\u4e0d\u5230 48GB \u7684\u200b\u200b VRAM \u4e2d\u5fae\u8c03\u5e76\u83b7\u5f97\u5feb\u901f\u63a8\u7406\u7684\u597d\u5904\uff01\u8981\u4f7f\u7528\u5feb\u901f\u63a8\u7406\uff0c\u9996\u5148\u5b89\u88c5 vllm\uff0c\u5e76\u4f7f\u7528 fast_inference \u5b9e\u4f8b\u5316 Unsloth\uff1a<\/p>\n<pre><code>pip install unsloth vllm\nfrom unsloth import FastLanguageModel\nmodel, tokenizer = FastLanguageModel.from_pretrained(\n    model_name = \"unsloth\/Llama-3.2-3B-Instruct\",\n    fast_inference = True,\n)\nmodel.fast_generate([\"Hello!\"])<\/code><\/pre>\n<h2>5\u3001Unsloth \u4e2d\u7684 vLLM \u53d1\u73b0<\/h2>\n<ul>\n<li>vLLM \u73b0\u5728\u53ef\u4ee5\u52a0\u8f7d Unsloth \u52a8\u6001 4 \u4f4d\u91cf\u5316\u3002\u5c31\u50cf\u6211\u4eec\u7684 1.58 \u4f4d\u52a8\u6001 R1 GGUF \u4e00\u6837\uff0c\u6211\u4eec\u8868\u660e\uff0c\u5c06\u67d0\u4e9b\u5c42\u52a8\u6001\u91cf\u5316\u4e3a 4 \u4f4d\uff0c\u5c06\u67d0\u4e9b\u5c42\u91cf\u5316\u4e3a 16 \u4f4d\u53ef\u4ee5\u663e\u8457\u63d0\u9ad8\u51c6\u786e\u6027\uff0c\u540c\u65f6\u4fdd\u6301\u6a21\u578b\u8f83\u5c0f\u3002<\/li>\n<li>\u6211\u4eec\u81ea\u52a8\u9009\u62e9\u591a\u4e2a\u53c2\u6570\u6765\u8003\u8651 RAM\u3001VRAM \u6548\u7387\u548c\u6700\u5927\u541e\u5410\u91cf\uff08\u4f8b\u5982\u5206\u5757\u9884\u586b\u5145\u4ee4\u724c\u7684\u6570\u91cf\u3001\u6700\u5927\u5e8f\u5217\u6570\u91cf\u7b49\uff09\u3002\u6211\u4eec\u9ed8\u8ba4\u5728 vLLM \u4e2d\u542f\u7528 -O3 \u5e76\u542f\u7528\u524d\u7f00\u7f13\u5b58\u3002\u6211\u4eec\u53d1\u73b0\u65e7 GPU \u4e0a\u7684 Flashinfer \u5b9e\u9645\u4e0a\u6162\u4e86 10%\u3002FP8 KV \u7f13\u5b58\u4f7f\u901f\u5ea6\u6162\u4e86 10%\uff0c\u4f46\u541e\u5410\u91cf\u6f5c\u529b\u589e\u52a0\u4e86\u4e00\u500d\u3002<\/li>\n<li>\u6211\u4eec\u5141\u8bb8\u901a\u8fc7\u89e3\u6790\u72b6\u6001\u5b57\u5178\u800c\u4e0d\u662f\u4ece\u78c1\u76d8\u52a0\u8f7d\u5728 vLLM \u4e2d\u52a0\u8f7d LoRA &#8211; \u8fd9\u53ef\u4ee5\u4f7f\u60a8\u7684 GRPO \u8bad\u7ec3\u8fd0\u884c\u901f\u5ea6\u63d0\u9ad8 1.5 \u500d\u3002\u4e00\u4e2a\u6d3b\u8dc3\u7684\u7814\u7a76\u9886\u57df\u662f\u4ee5\u67d0\u79cd\u65b9\u5f0f\u76f4\u63a5\u7f16\u8f91 vLLM \u4e2d\u7684 LoRA \u9002\u914d\u5668\uff08\u6211\u8fd8\u4e0d\u786e\u5b9a\u5982\u4f55\u64cd\u4f5c\uff09\u3002\u8fd9\u53ef\u4ee5\u5927\u5927\u63d0\u9ad8\u901f\u5ea6\uff0c\u56e0\u4e3a\u6211\u4eec\u73b0\u5728\u6b63\u5728\u8fdb\u884c\u4e0d\u5fc5\u8981\u7684 GPU \u6570\u636e\u79fb\u52a8\u3002<\/li>\n<li>vLLM \u4f1a\u51fa\u73b0\u5947\u602a\u7684\u968f\u673a VRAM \u5cf0\u503c\uff0c\u5c24\u5176\u662f\u5728\u6279\u91cf\u751f\u6210\u671f\u95f4\u3002\u6211\u4eec\u6dfb\u52a0\u4e86\u6279\u91cf\u751f\u6210\u529f\u80fd\u4ee5\u51cf\u5c11\u5185\u5b58\u5cf0\u503c\u3002<\/li>\n<\/ul>\n<hr>\n","protected":false},"excerpt":{"rendered":"<p>\u4eca\u5929\uff0c\u6211\u4eec\u5f88\u9ad8\u5174\u5728 Unsloth \u4e2d\u5f15\u5165\u63a8\u7406\u529f\u80fd\uff01DeepSeek \u7684 R1 \u7814\u7a76\u63ed\u793a\u4e86\u4e00\u4e2a\u201c\u987f\u609f\u65f6\u523b\u201d\uff0c\u5176\u4e2d R1-Zero \u901a\u8fc7\u4f7f\u7528\u7fa4\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316 (GRPO) \u81ea\u4e3b\u5b66\u4e60\u5206\u914d\u66f4\u591a\u601d\u8003\u65f6\u95f4\u800c\u65e0\u9700\u4eba\u5de5\u53cd\u9988\u3002 \u6211\u4eec\u589e\u5f3a\u4e86\u6574\u4e2a GRPO \u6d41\u7a0b\uff0c\u4f7f\u5176\u4f7f\u7528\u7684 VRAM \u6bd4 Hugging Face + FA2 \u5c11 80%\u3002\u8fd9\u6837\u4f60\u5c31\u53ef\u4ee5\u4f7f\u7528 Qwen2.5 (1.5B) \u5728\u4ec5 7GB \u7684 VRAM \u4e0a\u91cd\u73b0 R1-Zero \u7684\u201c\u987f\u609f\u65f6\u523b\u201d\u3002 \u8bd5\u7528\u6211\u4eec\u7684\u514d\u8d39 G\u200b\u200bRPO \u7b14\u8bb0\u672c\uff1a \u3002\u5bf9\u4e8e\u5177\u6709\u5176\u4ed6\u6a21\u578b\uff08\u5982 Phi-4\uff09\u7684 GRPO \u7b14\u8bb0\u672c\uff0c\u8bf7\u8bbf\u95ee 1\u3001\u4e3b\u8981\u7ec6\u8282 \u501f\u52a9 15GB VRAM\uff0cUnsloth \u53ef\u8ba9\u4f60\u5c06\u4efb\u4f55\u9ad8\u8fbe 15B \u53c2\u6570\u7684\u6a21\u578b\uff08\u5982 Llama 3.1 (8B)\u3001Phi-4 (14B)\u3001Mistral (7B) \u6216 Qwen2.5 (7B)\uff09\u8f6c\u6362\u4e3a\u63a8\u7406\u6a21\u578b \u6700\u4f4e\u8981\u6c42\uff1a\u4ec5 7GB VRAM [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[13],"tags":[],"class_list":["post-53816","post","type-post","status-publish","format-standard","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53816","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/comments?post=53816"}],"version-history":[{"count":0,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53816\/revisions"}],"wp:attachment":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/media?parent=53816"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/categories?post=53816"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/tags?post=53816"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}