{"id":53791,"date":"2025-02-16T11:50:39","date_gmt":"2025-02-16T03:50:39","guid":{"rendered":"https:\/\/fwq.ai\/blog\/53791\/"},"modified":"2025-02-16T11:50:39","modified_gmt":"2025-02-16T03:50:39","slug":"vptq%e4%bd%8e%e4%bd%8dllm%e9%87%8f%e5%8c%96%e7%ae%97%e6%b3%95","status":"publish","type":"post","link":"https:\/\/fwq.ai\/blog\/53791\/","title":{"rendered":"VPTQ\u4f4e\u4f4dLLM\u91cf\u5316\u7b97\u6cd5"},"content":{"rendered":"<p>LLM \u4f4e\u4f4d\u91cf\u5316\u7684\u6700\u65b0\u53d1\u5c55\uff0c\u4f8b\u5982 AQLM \u548c AutoRound\uff0c\u73b0\u5728\u5728\u4e0b\u6e38\u4efb\u52a1\u4e2d\u663e\u793a\u51fa\u53ef\u63a5\u53d7\u7684\u9000\u5316\u6c34\u5e73\uff0c\u5c24\u5176\u662f\u5bf9\u4e8e\u5927\u578b\u6a21\u578b\u3002 \u8bdd\u867d\u5982\u6b64\uff0c2 \u4f4d\u91cf\u5316\u5728\u5927\u591a\u6570\u60c5\u51b5\u4e0b\u4ecd\u4f1a\u5bfc\u81f4\u660e\u663e\u7684\u51c6\u786e\u6027\u635f\u5931\u3002<\/p>\n<p>\u4e00\u79cd\u5f88\u6709\u524d\u9014\u7684\u4f4e\u4f4d\u91cf\u5316\u7b97\u6cd5\u662f\u5fae\u8f6f\u63d0\u51fa\u7684 \uff08MIT \u8bb8\u53ef\u8bc1\uff09\u3002\u5b83\u4e8e 2024 \u5e74 10 \u6708\u63a8\u51fa\uff0c\u6b64\u540e\u5728\u91cf\u5316\u5927\u578b\u6a21\u578b\u65b9\u9762\u8868\u73b0\u51fa\u8272\uff0c\u6548\u7387\u6781\u9ad8\u3002<\/p>\n<p>\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u5c06\uff1a<\/p>\n<ul>\n<li>\u56de\u987e VPTQ \u91cf\u5316\u7b97\u6cd5<\/li>\n<li>\u6f14\u793a\u5982\u4f55\u4f7f\u7528 VPTQ \u6a21\u578b\uff0c\u5176\u4e2d\u8bb8\u591a\u6a21\u578b\u5df2\u7ecf\u53ef\u7528\u3002\u4f8b\u5982\uff0c\u6211\u4eec\u53ef\u4ee5\u8f7b\u677e\u627e\u5230 Llama 3.3 70B\u3001Llama 3.1 405B \u548c Qwen2.5 72B \u7684\u4f4e\u4f4d\u53d8\u4f53<\/li>\n<li>\u8bc4\u4f30\u8fd9\u4e9b\u6a21\u578b\u5e76\u8ba8\u8bba\u7ed3\u679c\uff0c\u4ee5\u4e86\u89e3 VPTQ \u6a21\u578b\u4f55\u65f6\u53ef\u4ee5\u6210\u4e3a\u751f\u4ea7\u4e2d LLM \u7684\u826f\u597d\u9009\u62e9\u3002<\/li>\n<\/ul>\n<p>\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5728 MMLU \u7b49\u4efb\u52a1\u4e0a\uff0c\u4f7f\u7528 VPTQ \u7684 2 \u4f4d\u91cf\u5316\u51e0\u4e4e\u5b9e\u73b0\u4e86\u4e0e\u539f\u59cb 16 \u4f4d\u6a21\u578b\u76f8\u5f53\u7684\u6027\u80fd\u3002\u6b64\u5916\uff0c\u5b83\u80fd\u591f\u5728\u5355\u4e2a GPU \u4e0a\u8fd0\u884c Llama 3.1 405B\uff0c\u540c\u65f6\u4f7f\u7528\u7684\u5185\u5b58\u6bd4 70B \u6a21\u578b\u5c11\uff01<\/p>\n<p>\u672c\u6587\u89e3\u91ca\u4e86\u8fd0\u884c VPTQ \u6a21\u578b\u548c\u8bc4\u4f30\u7684\u6240\u6709\u6b65\u9aa4\uff0c\u5e76\u5728\u4e2d\u5b9e\u73b0\u3002<\/p>\n<h2>1\u3001\u5411\u91cf\u8bad\u7ec3\u540e\u91cf\u5316<\/h2>\n<p>\u8fd9\u7bc7\u8bba\u6587\u4ecb\u7ecd\u4e86 VPTQ\uff1a \u3002\u5b83\u5229\u7528\u5411\u91cf\u91cf\u5316 (VQ)\uff0c\u8fd9\u662f\u4e00\u79cd\u5c06\u6743\u91cd\u7ec4\u8868\u793a\u4e3a\u5411\u91cf\u800c\u4e0d\u662f\u5355\u4e2a\u6807\u91cf\u7684\u6280\u672f\u3002<\/p>\n<p>\u6838\u5fc3\u601d\u60f3\u662f\u5c06 LLM \u7684\u6743\u91cd\u77e9\u9635\u91cd\u5851\u4e3a\u8f83\u5c0f\u7684\u5411\u91cf\u3002\u7136\u540e\u5c06\u6bcf\u4e2a\u5411\u91cf\u6620\u5c04\u5230\u9884\u5b9a\u4e49\u7801\u672c\u4e2d\u7684\u6700\u8fd1\u8d28\u5fc3\u3002\u6ce8\u610f\uff1a\u7801\u672c\u662f\u4e00\u7ec4\u53ef\u5b66\u4e60\u7684\u5019\u9009\u5411\u91cf\uff0c\u53ef\u7528\u4e8e\u5bf9\u6570\u636e\u8fdb\u884c\u7f16\u7801\u3002<\/p>\n<p>\u6b64\u6620\u5c04\u6700\u5c0f\u5316\u4e86\u8d28\u5fc3\u548c\u5411\u91cf\u4e4b\u95f4\u7684\u6b27\u51e0\u91cc\u5f97\u8ddd\u79bb\uff0c\u7528\u6307\u5411\u8d28\u5fc3\u7684\u7d22\u5f15\u66ff\u6362\u5411\u91cf\u3002\u8fd9\u79cd\u8f6c\u6362\u5141\u8bb8\u663e\u7740\u538b\u7f29\uff0c\u540c\u65f6\u4fdd\u6301\u51c6\u786e\u6027\u3002<\/p>\n<p>\u91cf\u5316\u8fc7\u7a0b\u7531\u4e8c\u9636\u4f18\u5316\u6846\u67b6\u6307\u5bfc\u3002\u5728\u8bef\u5dee\u5bf9\u6a21\u578b\u6027\u80fd\u7684\u5f71\u54cd\u5c3d\u53ef\u80fd\u5c0f\u7684\u7ea6\u675f\u4e0b\uff0c\u6700\u5c0f\u5316\u91cf\u5316\u8bef\u5dee\u3002Hessian \u77e9\u9635\u8868\u793a\u635f\u5931\u5bf9\u6743\u91cd\u53d8\u5316\u7684\u4e8c\u9636\u7075\u654f\u5ea6\u3002\u8fd9\u4e0e GPTQ \u7684\u4f5c\u7528\u7c7b\u4f3c\u3002<\/p>\n<p>\u5728 VPTQ \u7684\u8d85\u53c2\u6570\u4e2d\uff0c\u6211\u4eec\u6709\uff1a<\/p>\n<ul>\n<li>\u5411\u91cf\u957f\u5ea6 (v)<\/li>\n<li>\u8d28\u5fc3\u6570\u91cf (k)<\/li>\n<\/ul>\n<p>v \u548c k \u63a7\u5236\u51c6\u786e\u5ea6\u548c\u538b\u7f29\u7387\u4e4b\u95f4\u7684\u6743\u8861\u3002\u4f8b\u5982\uff0c\u8f83\u957f\u7684\u5411\u91cf\u4f1a\u51cf\u5c11\u6240\u9700\u7684\u8d28\u5fc3\u6570\u91cf\uff0c\u4ece\u800c\u63d0\u9ad8\u5185\u5b58\u6548\u7387\uff0c\u4f46\u53ef\u80fd\u4f1a\u589e\u52a0\u53cd\u91cf\u5316\u671f\u95f4\u7684\u8ba1\u7b97\u6210\u672c\u3002\u7801\u672c\u7684\u5927\u5c0f\u7531 k \u51b3\u5b9a\uff0c\u5176\u4e2d\u8f83\u5927\u7684 k \u53ef\u4ee5\u66f4\u597d\u5730\u8868\u793a\u6743\u91cd\u5206\u5e03\uff0c\u4f46\u4f1a\u6d88\u8017\u66f4\u591a\u5185\u5b58\u6765\u5b58\u50a8\u8d28\u5fc3\u3002<\/p>\n<p>\u6b8b\u5dee\u77e2\u91cf\u91cf\u5316 (RVQ) \u8fdb\u4e00\u6b65\u5b8c\u5584\u4e86\u8be5\u8fc7\u7a0b\u3002\u8fd9\u79cd\u591a\u9636\u6bb5\u6539\u8fdb\u4f7f\u7528\u8f85\u52a9\u7801\u672c\u6765\u6700\u5c0f\u5316\u6b8b\u5dee\u8bef\u5dee\uff0c\u4ece\u800c\u4ee5\u6700\u5c0f\u7684\u4f4d\u5f00\u9500\u5b9e\u73b0\u9ad8\u7cbe\u5ea6\u3002\u53e6\u4e00\u79cd\u6539\u8fdb\u89e3\u51b3\u4e86\u5f02\u5e38\u503c\u95ee\u9898\uff0c\u8fd9\u4e9b\u5f02\u5e38\u503c\u662f\u7f55\u89c1\u4f46\u4e0d\u6210\u6bd4\u4f8b\u7684\u5927\u6743\u91cd\uff0c\u53ef\u80fd\u4f1a\u626d\u66f2\u91cf\u5316\u7cbe\u5ea6\u3002\u4f7f\u7528\u4e13\u7528\u7801\u672c\u5355\u72ec\u5904\u7406\u5f02\u5e38\u503c\uff0c\u4ee5\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u5b83\u4eec\u5bf9\u6574\u4f53\u8bef\u5dee\u7684\u5f71\u54cd\u3002<\/p>\n<p>\u8d28\u5fc3\u7684\u521d\u59cb\u5316\u81f3\u5173\u91cd\u8981\uff0c\u4f7f\u7528 Hessian \u52a0\u6743 K \u5747\u503c\u805a\u7c7b\u5b8c\u6210\u3002\u8003\u8651\u5230 Hessian \u5bf9\u89d2\u7ebf\u6307\u793a\u7684\u6743\u91cd\u7684\u76f8\u5bf9\u91cd\u8981\u6027\uff0c\u8fd9\u79cd\u65b9\u6cd5\u53ef\u786e\u4fdd\u8d28\u5fc3\u4e0e\u6743\u91cd\u5206\u5e03\u5f88\u597d\u5730\u5bf9\u9f50\u3002\u4e0e\u7b80\u5355\u805a\u7c7b\u76f8\u6bd4\uff0c\u8fd9\u79cd\u52a0\u6743\u521d\u59cb\u5316\u53ef\u663e\u8457\u51cf\u5c11\u91cf\u5316\u8bef\u5dee\u3002<\/p>\n<p>\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0c\u6a21\u578b\u901a\u8fc7\u57fa\u4e8e\u7d22\u5f15\u4ece\u7801\u672c\u4e2d\u67e5\u627e\u8d28\u5fc3\u6765\u91cd\u5efa\u6743\u91cd\uff0c\u5982\u679c\u4f7f\u7528 RVQ\uff0c\u5219\u7ed3\u5408\u6b8b\u5dee\u6821\u6b63\u3002\u8fd9\u4f7f\u5f97\u63a8\u7406\u8fc7\u7a0b\u53d8\u5f97\u8f7b\u91cf\u7ea7\uff0c\u56e0\u4e3a\u5b83\u53ea\u6d89\u53ca\u7b80\u5355\u7684\u67e5\u627e\u548c\u6dfb\u52a0\u3002<\/p>\n<h2>2\u3001\u4f30\u8ba1 VPTQ \u6a21\u578b\u7684\u5e73\u5747\u4f4d\u5bbd<\/h2>\n<p>Microsoft \u63d0\u51fa\u4e86\u4e00\u79cd\u3002\u4ee5\u4e0b\u662f\u4ed6\u4eec\u7684\u505a\u6cd5\u3002<\/p>\n<p>\u5fae\u8f6f\u53d1\u5e03\u7684 VPTQ \u6a21\u578b\u4f7f\u7528\u7684\u6a21\u578b\u547d\u540d\u7ea6\u5b9a\u5305\u62ec\u6709\u5173\u5411\u91cf\u957f\u5ea6 (v)\u3001\u7801\u672c\u5927\u5c0f (k) \u548c\u6b8b\u5dee\u7801\u672c\u5927\u5c0f\u7684\u8be6\u7ec6\u4fe1\u606f\u3002\u4f8b\u5982\uff0c\u540d\u79f0\u201cMeta-Llama-3.1\u201370B-Instruct-v8-k65536\u2013256-woft\u201d\u5bf9\u5e94\u4e8e\u5177\u6709\u4ee5\u4e0b\u53c2\u6570\u7684\u6a21\u578b\u201cMeta-Llama-3.1\u201370B-Instruct\u201d\uff1a<\/p>\n<ul>\n<li>\u5411\u91cf\u957f\u5ea6\uff1av=8<\/li>\n<li>\u8d28\u5fc3\u6570\uff1ak=65536<\/li>\n<li>\u6b8b\u5dee\u8d28\u5fc3\u6570\uff1ak(res)=256<\/li>\n<\/ul>\n<p>\u6a21\u578b\u7684\u7b49\u6548\u4f4d\u5bbd\u53ef\u4ee5\u6309\u4ee5\u4e0b\u65b9\u5f0f\u8ba1\u7b97\uff1a<\/p>\n<ul>\n<li>\u7d22\u5f15\u4f4d\u5bbd\uff1a\u6bcf\u4e2a\u5411\u91cf\u7531\u8d28\u5fc3\u7d22\u5f15\u8868\u793a\u3002\u5bf9\u4e8e k=65536\uff0c\u6211\u4eec\u6709\u7d22\u5f15\u4f4d\u5bbd\uff1alog\u20612(65536) = 16 \u4f4d\u3002\u9664\u4ee5\u5411\u91cf\u957f\u5ea6 (v=8) \u53ef\u5f97\u51fa\uff1a16\/8= \u6bcf\u4e2a\u6743\u91cd 2 \u4f4d\u3002<\/li>\n<li>\u6b8b\u5dee\u7d22\u5f15\u4f4d\u5bbd\uff1a\u5bf9\u4e8e\u6b8b\u5dee\u8d28\u5fc3\uff0ck(res) = 256\uff0c\u7d22\u5f15\u4f4d\u5bbd\u4e3a\uff1alog2(256) = 8 \u4f4d\u3002\u9664\u4ee5\u5411\u91cf\u957f\u5ea6 (v=8) \u53ef\u5f97\u51fa\uff1a8\/8=\u6bcf\u4e2a\u6743\u91cd 1 \u4f4d\u3002<\/li>\n<li>\u603b\u4f4d\u5bbd\uff1a\u7ec4\u5408\u4f4d\u5bbd\u4e3a\uff1a2+1=\u6bcf\u4e2a\u6743\u91cd 3 \u4f4d\u3002<\/li>\n<\/ul>\n<p>\u8981\u4f30\u8ba1\u6a21\u578b\u5927\u5c0f\uff0c\u8bf7\u5c06\u53c2\u6570\u603b\u6570\u4e58\u4ee5\u4f4d\u5bbd\u5e76\u5c06\u5176\u8f6c\u6362\u4e3a\u5b57\u8282\u3002<\/p>\n<p>\u6ce8\u610f\uff1a\u6b64\u4f30\u8ba1\u76f8\u5f53\u51c6\u786e\uff0c\u4f46\u4e0d\u5305\u62ec\u7801\u672c\uff08\u67e5\u627e\u8868\uff09\u7684\u5927\u5c0f\u3001\u989d\u5916\u7684\u53c2\u6570\u5f00\u9500\u548c\u7528\u4e8e\u5b58\u50a8\u7d22\u5f15\u7684\u586b\u5145\u5f00\u9500\u3002<\/p>\n<h2>3\u3001\u8fd0\u884c VPTQ \u6a21\u578b<\/h2>\n<p>VPTQ \u96c6\u6210\u5230 Hugging Face Transformers \u4e2d\u3002\u8981\u8fd0\u884c\u6a21\u578b\uff0c\u4f60\u9700\u8981\u5b89\u88c5\u4ee5\u4e0b\u5185\u5bb9\uff1a<\/p>\n<pre><code>pip install --upgrade transformers vptq<\/code><\/pre>\n<p>Transformers \u4f1a\u81ea\u52a8\u68c0\u6d4b\u8be5\u6a21\u578b\u662f VPTQ \u6a21\u578b\u3002\u6211\u4eec\u65e0\u9700\u6267\u884c\u4efb\u4f55\u7279\u6b8a\u64cd\u4f5c\u5373\u53ef\u8fd0\u884c\u6a21\u578b\u3002\u5c31\u8fd9\u4e48\u7b80\u5355\uff1a<\/p>\n<pre><code>from transformers import AutoModelForCausalLM, AutoTokenizer\n\n# Load VPTQ-quantized model directly from HuggingFace Hub\nmodel = AutoModelForCausalLM.from_pretrained(\"VPTQ-community\/Meta-Llama-3.1-8B-Instruct-v8-k65536-256-woft\", device_map=\"auto\")\ntokenizer = AutoTokenizer.from_pretrained(\"VPTQ-community\/Meta-Llama-3.1-8B-Instruct-v8-k65536-256-woft\")\n# Simple inference\nprompt = \"Explain: Do not go gentle into that good night.\"\noutput = model.generate(**tokenizer(prompt, return_tensors=\"pt\").to(model.device))\nprint(tokenizer.decode(output[0], skip_special_tokens=True))<\/code><\/pre>\n<p>Hugging Face \u4e0a\u7684\u53d1\u5e03\u4e86\u8bb8\u591a\u6a21\u578b\u3002\u5728\u4e0b\u4e00\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u8bc4\u4f30\u4ee5\u4e0b\u6a21\u578b\u4e00\uff1a<\/p>\n<ul>\n<li>VPTQ-community\/Qwen2.5\u201372B-Instruct-v16-k65536\u201365536-woft<\/li>\n<li>VPTQ-community\/Qwen2.5\u201372B-Instruct-v8-k65536\u20130-woft<\/li>\n<li>VPTQ-community\/Qwen2.5\u201372B-Instruct-v16-k65536\u201332768-woft<\/li>\n<li>VPTQ-community\/Meta-Llama-3.3\u201370B-Instruct-v8-k65536\u20130-woft<\/li>\n<li>VPTQ-community\/Meta-Llama-3.3\u201370B-Instruct-v16-k65536\u20131024-woft<\/li>\n<li>VPTQ-community\/ Meta-Llama-3.1\u2013405B-Instruct-v16-k65536\u201365536-woft<\/li>\n<li>VPTQ-community\/Meta-Llama-3.1\u2013405B-Instruct-v16-k32768\u201332768-woft<\/li>\n<li>VPTQ-community\/Meta-Llama-3.1\u2013405B-Instruct-v16-k65536\u20131024-woft<\/li>\n<li>VPTQ-community\/Meta-Llama-3.1\u2013405B-Instruct-v16-k65536\u2013256-woft<\/li>\n<li>VPTQ-community\/Meta-Llama-3.1\u2013405B-Instruct-v16-k65536\u201364-woft<\/li>\n<\/ul>\n<p>\u6ce8\u610f\uff1a\u8fd9\u4e9b Qwen \u548c Llama \u6a21\u578b\u5206\u522b\u4f7f\u7528 llama \u548c qwen \u8bb8\u53ef\u8bc1\u53d1\u5e03\u3002<\/p>\n<p>\u6240\u6709\u8fd9\u4e9b Llama 3.3 \u548c Qwen2.5 \u6a21\u578b\u90fd\u53ef\u4ee5\u5728\u5355\u4e2a 24 GB GPU \u4e0a\u8fd0\u884c\uff01Llama 3.1 405B \u6a21\u578b\u867d\u7136\u6bd4\u539f\u59cb\u6a21\u578b\u5c0f\u5f97\u591a\uff0c\u4f46\u4ecd\u7136\u9700\u8981\u5927\u91cf GPU \u5185\u5b58\uff0c\u4f46\u6211\u53ef\u4ee5\u4f7f\u7528 RunPod \u63d0\u4f9b\u7684 H200 SXM\uff08\u63a8\u8350\u94fe\u63a5\uff09\u8fd0\u884c\u6240\u6709\u6a21\u578b\uff0c\u56e0\u6b64\u5b83\u4eec\u53ef\u4ee5\u5728\u5355\u4e2a GPU \u4e0a\u8fd0\u884c\u3002<\/p>\n<p>\u6211\u8fd8\u6ca1\u6709\u6d4b\u8bd5\u8fc7\uff0c\u4f46\u5e94\u8be5\u4e5f\u53ef\u4ee5\u4f7f\u7528 QLoRA \u65b9\u6cd5\u5728 VPTQ \u6a21\u578b\u4e4b\u4e0a\u5fae\u8c03 LoRA \u9002\u914d\u5668\u3002\u6700\u7cdf\u7cd5\u7684\u662f\uff0c\u5728 Hugging Face PEFT \u5e93\u4e2d\u6dfb\u52a0 VPTQ \u652f\u6301\u7684\u4ee3\u7801\u53ea\u6709\u51e0\u884c\u3002<\/p>\n<h2>4\u3001\u8bc4\u4f30 VPTQ \u6a21\u578b<\/h2>\n<p>\u4e3a\u4e86\u8bc4\u4f30\u91cf\u5316\u7b97\u6cd5\u7684\u597d\u574f\uff0c\u6211\u901a\u5e38\u4f9d\u8d56 MMLU\u3002\u6211\u8ba1\u7b97\u91cf\u5316\u6a21\u578b\u7684\u51c6\u786e\u6027\u5e76\u5c06\u5176\u4e0e\u539f\u59cb\u6a21\u578b\u7684\u51c6\u786e\u6027\u8fdb\u884c\u6bd4\u8f83\u3002\u5982\u679c\u4e24\u79cd\u7cbe\u5ea6\u63a5\u8fd1\uff0c\u5219\u91cf\u5316\u6548\u679c\u826f\u597d\u3002<\/p>\n<p>\u4f46\u8bf7\u6ce8\u610f\uff0c\u8fd9\u8fdc\u8fdc\u4e0d\u8db3\u4ee5\u8bc4\u4f30\u91cf\u5316\u6a21\u578b\u3002\u91cf\u5316\u6a21\u578b\u53ef\u80fd\u5bf9\u67d0\u4e9b\u4efb\u52a1\u51c6\u786e\uff0c\u4f46\u5bf9\u5176\u4ed6\u4efb\u52a1\u5b8c\u5168\u65e0\u6548\u3002\u4f8b\u5982\uff0c\u4f7f\u7528 AutoRound \u91cf\u5316\uff0c\u6211\u770b\u5230 2 \u4f4d\u6a21\u578b\u5728 MMLU \u4e0a\u8868\u73b0\u975e\u5e38\u597d\uff0c\u4f46\u751f\u6210\u4e71\u7801\u3002\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u6211\u5efa\u8bae\u4e5f\u5728\u67d0\u4e9b\u751f\u6210\u4efb\u52a1\uff08\u5982 IFEval \u6216 MATH\uff09\u4e0a\u8bc4\u4f30\u6a21\u578b\uff0c\u4ee5\u786e\u4fdd\u6a21\u578b\u6ca1\u6709\u635f\u574f\u3002<\/p>\n<p>\u5bf9\u4e8e\u672c\u6587\uff0c\u6211\u4ec5\u4f7f\u7528 MMLU \u8bc4\u4f30\u6240\u6709\u6a21\u578b\uff0c\u5e76\u4f7f\u7528 MMLU-PRO\u3001MuSR \u548c GPQA \u8bc4\u4f30 Llama 3.3 \u548c Qwen2.5\u3002\u6211\u6ca1\u6709\u8fd0\u884c\u751f\u6210\u57fa\u51c6\u6d4b\u8bd5\uff0c\u5e76\u4e14\u4ec5\u5728 MMLU \u4e0a\u8bc4\u4f30\u4e86 Llama 3.1 405B\uff0c\u56e0\u4e3a\u8bc4\u4f30\u6210\u672c\u5f88\u9ad8\u3002\u6211\u8ba1\u5212\u5728 vLLM \u4e2d\u6dfb\u52a0 VPTQ \u652f\u6301\u65f6\u8fd9\u6837\u505a\u3002<\/p>\n<p>\u6211\u4f7f\u7528\u8bc4\u4f30\u5de5\u5177\u6765\u8fd0\u884c\u6b64\u8bc4\u4f30\u3002<\/p>\n<p>  \u6bcf\u4e2a\u6570\u636e\u70b9\u90fd\u6807\u6709\u6a21\u578b\u7684\u4f30\u8ba1\u5e73\u5747\u4f4d\u5bbd\u3002ChatGPT \u4f7f\u7528\u751f\u6210\u7684 Python \u751f\u6210\u56fe\u8868\uff0c\u5229\u7528\u6211\u7684\u8bc4\u4f30\u6570\u636e\u3002 <\/p>\n<p>\u8fd9\u4e9b\u7ed3\u679c\u975e\u5e38\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u3002<\/p>\n<p>\u4f5c\u4e3a\u53c2\u8003\uff0c\u4f7f\u7528 AutoRound\uff0c\u5728\u76f8\u540c\u7684\u8bc4\u4f30\u8bbe\u7f6e\u4e0b\uff0c\u6211\u4f7f\u7528 2 \u4f4d\u91cf\u5316\u53ef\u4ee5\u5b9e\u73b0\u7684\u6700\u4f73\u51c6\u786e\u5ea6\u4e3a 73.71\u3002\u8fd9\u4e5f\u662f\u4f7f\u7528 Qwen2.5 72B Instruct \u5b9e\u73b0\u7684\u3002<\/p>\n<p>2 \u4f4d Qwen2.5 \u6a21\u578b\u7684\u6027\u80fd\u4ec5\u6bd4\u539f\u59cb\u6a21\u578b\u4f4e 5 \u5206\u3002\u6211\u4eec\u975e\u5e38\u63a5\u8fd1\u8ba9 2 \u4f4d\u6a21\u578b\u4e0e\u539f\u59cb 16 \u4f4d\u6a21\u578b\u4e00\u6837\u597d\u3002<\/p>\n<p>\u5f53\u6211\u5c1d\u8bd5\u4f7f\u7528 Llama 3.3 \u8fdb\u884c AutoRound 2 \u4f4d\u91cf\u5316\u65f6\uff0c\u5b83\u5b8c\u5168\u5931\u8d25\u4e86\u3002\u4f7f\u7528 VPTQ\uff0c\u5b83\u8fd0\u884c\u826f\u597d\uff0cMMLU \u51c6\u786e\u5ea6\u63a5\u8fd1 75.0\u3002<\/p>\n<p>\u6b63\u5982\u9884\u671f\u7684\u90a3\u6837\uff0c\u50cf Llama 3.1 405B \u8fd9\u6837\u7684\u8f83\u5927\u6a21\u578b\u5bf9\u4f4e\u4f4d\u91cf\u5316\u66f4\u5177\u9c81\u68d2\u6027\u3002\u5373\u4f7f\u5e73\u5747\u4f4d\u5bbd\u4e3a 1.5\uff0c\u8be5\u6a21\u578b\u4ecd\u53ef\u5b9e\u73b0\u63a5\u8fd1<\/p>\n<p>\u8bbe\u7f6e\u4e3a 70.0\uff0c\u8fd9\u4ec5\u7565\u4f4e\u4e8e\u6211\u8bc4\u4f30 Qwen2.5 7B Instruct \u65f6\u5f97\u5230\u7684\u7ed3\u679c\u3002\u7136\u800c\uff0c\u5728\u5b9e\u8df5\u4e2d\uff0c\u8fd9\u4e2a\u6a21\u578b\u5e76\u4e0d\u9002\u7528\u4e8e\u4efb\u4f55\u4e8b\u60c5\uff0c\u56e0\u4e3a\u5b83\u4ecd\u7136\u6bd4\u5176\u4ed6\u8868\u73b0\u66f4\u597d\u7684\u6a21\u578b\u5927\u5f97\u591a\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u6211\u8fd0\u884c\u4e86 MMLU-PRO\u3001MuSR \u548c GPQA \u6765\u8bc4\u4f30 Llama 3.3 \u548c Qwen2.5\uff1a<\/p>\n<p>\u6240\u6709\u8fd9\u4e9b\u5206\u6570\u5bf9\u4e8e 2 \u4f4d\u6a21\u578b\u6765\u8bf4\u90fd\u975e\u5e38\u51fa\u8272\u3002\u4f5c\u4e3a\u53c2\u8003\uff0c\u6839\u636e OpenLLM \u6392\u884c\u699c\uff0c\u539f\u59cb\uff0816 \u4f4d\uff09Qwen2.5 72B Instruct \u5728 MMLU-PRO \u4e0a\u7684\u6027\u80fd\u4e3a 51.4\u3002\u8fd9\u4e0e 2 \u4f4d VPTQ \u6a21\u578b\u7684\u6027\u80fd\u76f8\u5dee\u4e0d\u8fdc\u3002\u6ce8\u610f\uff1aOpenLLM \u6392\u884c\u699c\u5c06 MMLU-PRO \u4f5c\u4e3a 5 \u6b21\u4efb\u52a1\u8fd0\u884c\uff0c\u800c\u6211\u5c06\u5176\u4f5c\u4e3a 0 \u6b21\u4efb\u52a1\u8fd0\u884c\uff0c\u8fd9\u4f7f VPTQ \u6a21\u578b\u5904\u4e8e\u52a3\u52bf\u3002<\/p>\n<h2>5\u3001\u7ed3\u675f\u8bed<\/h2>\n<p>\u5bf9\u4e8e\u4f4e\u4f4d\u91cf\u5316\uff0cVPTQ \u7684\u51c6\u786e\u6027\u786e\u5b9e\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u3002\u6b64\u5916\uff0c\u4e0e\u4e4b\u524d\u5728\u4f4e\u7cbe\u5ea6\u4e0b\u53ef\u80fd\u5b8c\u5168\u7834\u574f\u6a21\u578b\u7684\u7b97\u6cd5\u4e0d\u540c\uff0cVPTQ \u4f3c\u4e4e\u66f4\u4e3a\u7a33\u5065\u3002\u6700\u540e\uff0c\u6211\u4eec\u5f97\u5230\u4e86\u4e00\u4e2a\u4e0d\u4f1a\u4ea7\u751f\u4e71\u7801\u7684 2 \u4f4d Llama 3.3 70B\u3002<\/p>\n<p>\u968f\u7740\u4f4e\u4f4d\u91cf\u5316\u8d8a\u6765\u8d8a\u597d\uff0c\u91cf\u5316\u8f83\u5927\u7684\u6a21\u578b\u53ef\u80fd\u6bd4\u4f7f\u7528\u8f83\u5c0f\u7684 16 \u4f4d\u6a21\u578b\u66f4\u53ef\u53d6\u3002\u4e3b\u8981\u7f3a\u70b9\u662f\u63a8\u7406\u901f\u5ea6\u4f1a\u6bd4\u8f83\u5c0f\u7684\u6a21\u578b\u6162\uff0c\u56e0\u4e3a\u5b83\u4eec\u6709\u66f4\u591a\u7684\u53c2\u6570\u3002<\/p>\n<p>\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4ec5\u63a2\u8ba8\u4e86 VPTQ \u4f5c\u8005\u6b63\u5f0f\u53d1\u5e03\u7684\u91cf\u5316\u6a21\u578b\u3002\u4f7f\u7528 VPTQ \u91cf\u5316\u81ea\u5df1\u7684\u6a21\u578b\u6709\u591a\u5bb9\u6613\uff1f<\/p>\n<p>\u8be5\u7b97\u6cd5\u4f3c\u4e4e\u4e0e AutoRound \u4e00\u6837\u6602\u8d35\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u5e94\u8be5\u53ef\u4ee5\u5728\u4e91\u7aef\u4ee5\u4e0d\u5230 10 \u7f8e\u5143\u7684\u4ef7\u683c\u91cf\u5316 70B \u6a21\u578b\u3002\u4f46\u662f\uff0c\u91cf\u5316\u7b97\u6cd5\u5c1a\u672a\u5728 GitHub \u5b58\u50a8\u5e93\u4e2d\u5b8c\u5168\u53d1\u5e03\u3002\u4f5c\u8005\u4f3c\u4e4e\u60f3\u5bf9\u5176\u8fdb\u884c\u8fdb\u4e00\u6b65\u5b8c\u5584\u3002\u636e\u6211\u6240\u77e5\uff0c\u5b83\u51e0\u4e4e\u5b8c\u5168\u5728 \u53ef\u7528\u3002<\/p>\n<p>\u8ba1\u7b97 Hessian \u77e9\u9635\u7684\u90e8\u5206\u4ecd\u7136\u7f3a\u5931\uff08\u622a\u81f3 2025 \u5e74 1 \u6708 25 \u65e5\uff09\u3002\u6211\u4eec\u5fc5\u987b\u4f9d\u9760\u9884\u5148\u8ba1\u7b97\u7684\u77e9\u9635\u6216\u4f7f\u7528\u7b2c\u4e09\u65b9\u8f6f\u4ef6\u8ba1\u7b97\u5b83\u4eec\u3002<\/p>\n<hr>\n","protected":false},"excerpt":{"rendered":"<p>LLM \u4f4e\u4f4d\u91cf\u5316\u7684\u6700\u65b0\u53d1\u5c55\uff0c\u4f8b\u5982 AQLM \u548c AutoRound\uff0c\u73b0\u5728\u5728\u4e0b\u6e38\u4efb\u52a1\u4e2d\u663e\u793a\u51fa\u53ef\u63a5\u53d7\u7684\u9000\u5316\u6c34\u5e73\uff0c\u5c24\u5176\u662f\u5bf9\u4e8e\u5927\u578b\u6a21\u578b\u3002 \u8bdd\u867d\u5982\u6b64\uff0c2 \u4f4d\u91cf\u5316\u5728\u5927\u591a\u6570\u60c5\u51b5\u4e0b\u4ecd\u4f1a\u5bfc\u81f4\u660e\u663e\u7684\u51c6\u786e\u6027\u635f\u5931\u3002 \u4e00\u79cd\u5f88\u6709\u524d\u9014\u7684\u4f4e\u4f4d\u91cf\u5316\u7b97\u6cd5\u662f\u5fae\u8f6f\u63d0\u51fa\u7684 \uff08MIT \u8bb8\u53ef\u8bc1\uff09\u3002\u5b83\u4e8e 2024 \u5e74 10 \u6708\u63a8\u51fa\uff0c\u6b64\u540e\u5728\u91cf\u5316\u5927\u578b\u6a21\u578b\u65b9\u9762\u8868\u73b0\u51fa\u8272\uff0c\u6548\u7387\u6781\u9ad8\u3002 \u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u5c06\uff1a \u56de\u987e VPTQ \u91cf\u5316\u7b97\u6cd5 \u6f14\u793a\u5982\u4f55\u4f7f\u7528 VPTQ \u6a21\u578b\uff0c\u5176\u4e2d\u8bb8\u591a\u6a21\u578b\u5df2\u7ecf\u53ef\u7528\u3002\u4f8b\u5982\uff0c\u6211\u4eec\u53ef\u4ee5\u8f7b\u677e\u627e\u5230 Llama 3.3 70B\u3001Llama 3.1 405B \u548c Qwen2.5 72B \u7684\u4f4e\u4f4d\u53d8\u4f53 \u8bc4\u4f30\u8fd9\u4e9b\u6a21\u578b\u5e76\u8ba8\u8bba\u7ed3\u679c\uff0c\u4ee5\u4e86\u89e3 VPTQ \u6a21\u578b\u4f55\u65f6\u53ef\u4ee5\u6210\u4e3a\u751f\u4ea7\u4e2d LLM \u7684\u826f\u597d\u9009\u62e9\u3002 \u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5728 MMLU \u7b49\u4efb\u52a1\u4e0a\uff0c\u4f7f\u7528 VPTQ \u7684 2 \u4f4d\u91cf\u5316\u51e0\u4e4e\u5b9e\u73b0\u4e86\u4e0e\u539f\u59cb 16 \u4f4d\u6a21\u578b\u76f8\u5f53\u7684\u6027\u80fd\u3002\u6b64\u5916\uff0c\u5b83\u80fd\u591f\u5728\u5355\u4e2a GPU \u4e0a\u8fd0\u884c Llama 3.1 405B\uff0c\u540c\u65f6\u4f7f\u7528\u7684\u5185\u5b58\u6bd4 70B \u6a21\u578b\u5c11\uff01 \u672c\u6587\u89e3\u91ca\u4e86\u8fd0\u884c VPTQ \u6a21\u578b\u548c\u8bc4\u4f30\u7684\u6240\u6709\u6b65\u9aa4\uff0c\u5e76\u5728\u4e2d\u5b9e\u73b0\u3002 1\u3001\u5411\u91cf\u8bad\u7ec3\u540e\u91cf\u5316 [&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-53791","post","type-post","status-publish","format-standard","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53791","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=53791"}],"version-history":[{"count":0,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53791\/revisions"}],"wp:attachment":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/media?parent=53791"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/categories?post=53791"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/tags?post=53791"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}