{"id":20794,"date":"2025-02-11T15:03:33","date_gmt":"2025-02-11T07:03:33","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=20794"},"modified":"2025-02-11T15:03:33","modified_gmt":"2025-02-11T07:03:33","slug":"jinxuyao-14gb-xianbao","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/ja\/jinxuyao-14gb-xianbao\/","title":{"rendered":"\u4ec5\u9700\u8981 14GB \u663e\u5b58\u672c\u5730\u8fd0\u884c DeepSeek-Coder V3\/R1 (Q4_K_M \u91cf\u5316)"},"content":{"rendered":"<h2>\u6458\u8981<\/h2>\n<blockquote><p><strong>2025 \u5e74 2 \u6708 10 \u65e5<\/strong>: \u5728\u5355 GPU (24GB \u663e\u5b58) \/ \u591a GPU \u548c 382GB \u5185\u5b58\u4e0a\u652f\u6301 DeepseekR1 \u548c V3\uff0c\u901f\u5ea6\u63d0\u5347\u9ad8\u8fbe 3~28 \u500d\u3002<\/p><\/blockquote>\n<p>\u5927\u5bb6\u597d\uff0cKTransformers \u56e2\u961f (\u524d\u8eab\u4e3a CPU\/GPU \u6df7\u5408\u63a8\u7406\u5f00\u6e90\u9879\u76ee\u56e2\u961f\uff0c\u4ee5 DeepSeek-V2 \u800c\u95fb\u540d) \u5411\u5927\u5bb6\u95ee\u597d\u3002<\/p>\n<p><a href=\"https:\/\/www.kdjingpai.com\/ktransformers\/\">KTransformers<\/a> \u56e2\u961f\u6536\u5230\u4e86\u5927\u5bb6\u5bf9 DeepSeek-R1\/V3 \u652f\u6301\u7684\u8bf7\u6c42\uff0c\u5e76\u4e14\u975e\u5e38\u6fc0\u52a8\u5730\u5ba3\u5e03\u7ec8\u4e8e\u4ea4\u4ed8\u4e86\uff01<br \/>\n\u5bf9\u4e8e\u6b64\u6b21\u7b49\u5f85\u6df1\u611f\u62b1\u6b49\uff0c\u4f46 KTransformers \u56e2\u961f\u4e00\u76f4\u5728\u915d\u917f\u4e00\u4e9b\u771f\u6b63\u4ee4\u4eba\u60ca\u53f9\u7684\u4e1c\u897f\uff01<\/p>\n<p>\u4eca\u5929\uff0cKTransformers \u56e2\u961f\u81ea\u8c6a\u5730\u5ba3\u5e03\uff0c\u4e0d\u4ec5\u652f\u6301 DeepSeek-R1\/V3\uff0c\u6b63\u5982\u4ee5\u4e0b\u89c6\u9891\u6240\u793a\uff1a<\/p>\n<p>https:\/\/github.com\/user-attachments\/assets\/ebd70bfa-b2c1-4abb-ae3b-296ed38aa285<\/p>\n<p>&nbsp;<\/p>\n<ul>\n<li><strong>[\u6700\u65b0!!!] \u672c\u5730 671B DeepSeek-Coder-V3\/R1:<\/strong>\u00a0\u4ec5\u4f7f\u7528 14GB \u663e\u5b58\u548c 382GB \u5185\u5b58\u8fd0\u884c\u5176 Q4_K_M \u7248\u672c\u3002\n<ul>\n<li>Prefill \u901f\u5ea6 (tokens\/s):\n<ul>\n<li>KTransfermor: 54.21 (32 \u6838) \u2192 74.362 (\u53cc\u8def\uff0c2\u00d732 \u6838) \u2192 255.26 (\u4f18\u5316\u7684\u57fa\u4e8e AMX \u7684 MoE \u5185\u6838\uff0c\u4ec5 V0.3) \u2192 286.55 (\u9009\u62e9\u6027\u4f7f\u7528 6 \u4e2a\u4e13\u5bb6\uff0c\u4ec5 V0.3)<\/li>\n<li>\u4e0e <a href=\"https:\/\/www.kdjingpai.com\/llamacpp\/\">llama.cpp<\/a> \u5728 2\u00d732 \u6838\u4e0b\u7684 10.31 tokens\/s \u76f8\u6bd4\uff0c\u5b9e\u73b0\u4e86\u9ad8\u8fbe\u00a0<strong>27.79 \u500d\u7684\u52a0\u901f<\/strong>\u3002<\/li>\n<\/ul>\n<\/li>\n<li>Decode \u901f\u5ea6 (tokens\/s):\n<ul>\n<li>KTransfermor: 8.73 (32 \u6838) \u2192 11.26 (\u53cc\u8def\uff0c2\u00d732 \u6838) \u2192 13.69 (\u9009\u62e9\u6027\u4f7f\u7528 6 \u4e2a\u4e13\u5bb6\uff0c\u4ec5 V0.3)<\/li>\n<li>\u4e0e llama.cpp \u5728 2\u00d732 \u6838\u4e0b\u7684 4.51 tokens\/s \u76f8\u6bd4\uff0c\u5b9e\u73b0\u4e86\u9ad8\u8fbe\u00a0<strong>3.03 \u500d\u7684\u52a0\u901f<\/strong>\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p>KTransformers \u56e2\u961f\u8fd8\u7ed9\u51fa\u4e86\u5373\u5c06\u5230\u6765\u7684\u4f18\u5316\u9884\u89c8\uff0c\u5305\u62ec Intel AMX \u52a0\u901f\u5185\u6838\u548c\u9009\u62e9\u6027\u4e13\u5bb6\u6fc0\u6d3b\u65b9\u6cd5\uff0c\u8fd9\u5c06\u663e\u8457\u63d0\u9ad8\u6027\u80fd\u3002\u501f\u52a9 V0.3-preview\uff0c\u9884\u586b\u5145 (prefill) \u901f\u5ea6\u9ad8\u8fbe 286 tokens\/s\uff0c\u6bd4\u672c\u5730\u63a8\u7406\u7684 llama.cpp \u5feb\u00a0<strong>28 \u500d<\/strong>\u3002<br \/>\n\u4e8c\u8fdb\u5236\u53d1\u884c\u7248\u73b0\u5df2\u53d1\u5e03\uff0c\u6e90\u4ee3\u7801\u5c06\u5c3d\u5feb\u53d1\u5e03\uff01<a href=\"https:\/\/github.com\/kvcache-ai\/ktransformers\/releases\/download\/v0.1.4\/ktransformers-0.3.0rc0+cu126torch26fancy-cp311-cp311-linux_x86_64.whl\">\u5728\u6b64\u5904\u67e5\u770b wheel \u5305<\/a>\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>\u51c6\u5907\u6761\u4ef6<\/h2>\n<p>KTransformers \u56e2\u961f\u5728\u4ee5\u4e0b\u914d\u7f6e\u4e0a\u8fd0\u884c\u4e86\u6700\u4f73\u6027\u80fd\u6d4b\u8bd5 (V0.2)\uff1a<\/p>\n<p>CPU: Intel (R) Xeon (R) Gold 6454S 1T \u5185\u5b58 (2 \u4e2a NUMA \u8282\u70b9)<\/p>\n<p>GPU: 4090D 24G \u663e\u5b58<\/p>\n<p>\u5185\u5b58: \u6807\u51c6 DDR5-4800 \u670d\u52a1\u5668\u5185\u5b58 (1 TB)<\/p>\n<p>&nbsp;<\/p>\n<h2>\u57fa\u51c6\u6d4b\u8bd5\u7ed3\u679c<\/h2>\n<h3>V0.2<\/h3>\n<h4>\u8bbe\u7f6e<\/h4>\n<ul>\n<li>\u6a21\u578b: DeepseekV3-q4km (int4)<\/li>\n<li>CPU: cpu_model_name: Intel (R) Xeon (R) Gold 6454S, \u6bcf\u8def 32 \u6838\uff0c2 \u8def\uff0c2 \u4e2a numa \u8282\u70b9<\/li>\n<li>GPU: 4090D 24G \u663e\u5b58<\/li>\n<li>KTransformers \u56e2\u961f\u5728\u5145\u5206\u9884\u70ed\u540e\u8fdb\u884c\u6d4b\u8bd5<\/li>\n<\/ul>\n<h4>\u5185\u5b58\u6d88\u8017:<\/h4>\n<ul>\n<li>\u5355\u8def: 382G \u5185\u5b58\uff0c\u81f3\u5c11 14GB \u663e\u5b58<\/li>\n<li>\u53cc\u8def: 1T \u5185\u5b58\uff0c\u81f3\u5c11 14GB \u663e\u5b58<\/li>\n<\/ul>\n<h4>\u57fa\u51c6\u6d4b\u8bd5\u7ed3\u679c<\/h4>\n<p>\u201c6 experts\u201d \u60c5\u51b5\u662f V0.3 \u9884\u89c8\u7248\u7684\u4e00\u90e8\u5206<\/p>\n<p>| Prompt<\/p>\n<table>\n<thead>\n<tr>\n<th>(500 <a href=\"https:\/\/www.kdjingpai.com\/tokenization\/\">tokens<\/a>)<\/th>\n<th>\u53cc\u8def Ktrans (6 experts)<\/th>\n<th>\u53cc\u8def Ktrans (8 experts)<\/th>\n<th>\u5355\u8def Ktrans (6 experts)<\/th>\n<th>\u5355\u8def Ktrans (8 experts)<\/th>\n<th>llama.cpp (8 experts)<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Prefill token\/s<\/td>\n<td>97.32<\/td>\n<td>82.94<\/td>\n<td>65.14<\/td>\n<td>54.21<\/td>\n<td>10.31<\/td>\n<\/tr>\n<tr>\n<td>Decode token\/s<\/td>\n<td>13.69<\/td>\n<td>12.208<\/td>\n<td>10.303<\/td>\n<td>8.73<\/td>\n<td>4.51<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>\u89e3\u7801\u901f\u5ea6\u6700\u9ad8\u63d0\u5347\u00a0<u>3.03 \u500d<\/u>\uff0c\u9884\u586b\u5145\u901f\u5ea6\u6700\u9ad8\u63d0\u5347\u00a0<u>9.44 \u500d<\/u>\u3002<\/strong>\u00a0\u770b\u8d77\u6765 KTransformers \u5728\u89e3\u7801\u52a0\u901f\u65b9\u9762\u63d0\u5347\u4e0d\u5982\u9884\u586b\u5145\u660e\u663e\uff0c\u89e3\u7801\u7684\u4f18\u5316\u7a7a\u95f4\u4ecd\u7136\u5f88\u5927\u3002<\/p>\n<h3>V0.3-Preview<\/h3>\n<h4>\u8bbe\u7f6e<\/h4>\n<ul>\n<li>\u6a21\u578b: DeepseekV3-BF16 (\u5728\u7ebf\u91cf\u5316\u4e3a int8 \u7528\u4e8e CPU\uff0cint4 \u7528\u4e8e GPU)<\/li>\n<li>CPU: cpu_model_name: Intel (R) Xeon (R) Gold 6454S, \u6bcf\u8def 32 \u6838\uff0c2 \u8def\uff0c2 \u4e2a numa \u8282\u70b9<\/li>\n<li>GPU: (1~4)x 4090D 24GVRAM (\u66f4\u957f\u7684 prompt \u9700\u8981\u66f4\u591a\u663e\u5b58)<\/li>\n<\/ul>\n<h4>\u5185\u5b58\u6d88\u8017:<\/h4>\n<ul>\n<li>644GB \u5185\u5b58\uff0c\u81f3\u5c11 14GB \u663e\u5b58<\/li>\n<\/ul>\n<h4>\u57fa\u51c6\u6d4b\u8bd5\u7ed3\u679c<\/h4>\n<table>\n<thead>\n<tr>\n<th>Prompt length<\/th>\n<th>1K<\/th>\n<th>2K<\/th>\n<th>4K<\/th>\n<th>8K<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>KTrans (8 experts) Prefill token\/s<\/td>\n<td>185.96<\/td>\n<td>255.26<\/td>\n<td>252.58<\/td>\n<td>195.62<\/td>\n<\/tr>\n<tr>\n<td>KTrans (6 experts) Prefill token\/s<\/td>\n<td>203.70<\/td>\n<td>286.55<\/td>\n<td>271.08<\/td>\n<td>207.20<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>KTrans V0.3 \u7684\u9884\u586b\u5145\u901f\u5ea6\u6bd4 KTrans V0.2 \u5feb\u00a0<u>3.45 \u500d<\/u>\uff0c\u6bd4 llama.cpp \u5feb\u00a0<u>27.79 \u500d<\/u>\u3002<\/strong>\u00a0\u8fd9\u4e2a\u9884\u586b\u5145\u901f\u5ea6\u7684\u63d0\u5347\u786e\u5b9e\u975e\u5e38\u60ca\u4eba\uff0c\u770b\u8d77\u6765 KTransformers \u5728\u9884\u586b\u5145\u4f18\u5316\u65b9\u9762\u4e0b\u4e86\u5f88\u5927\u529f\u592b\u3002<br \/>\n<strong>\u89e3\u7801\u901f\u5ea6\u4e0e KTrans V0.2 (6 experts \u7248\u672c) \u76f8\u540c\uff0c\u56e0\u6b64\u7701\u7565<\/strong>\u00a0\u770b\u6765 V0.3 \u7248\u672c\u4e3b\u8981\u5173\u6ce8\u5728\u9884\u586b\u5145\u901f\u5ea6\u4e0a\u7684\u63d0\u5347\u3002<\/p>\n<p>\u4e3b\u8981\u7684\u52a0\u901f\u6765\u81ea<\/p>\n<ul>\n<li>Intel AMX \u6307\u4ee4\u96c6\u548c KTransformers \u56e2\u961f\u4e13\u95e8\u8bbe\u8ba1\u7684\u7f13\u5b58\u53cb\u597d\u578b\u5185\u5b58\u5e03\u5c40<\/li>\n<li>\u57fa\u4e8e\u9886\u57df\u5916\u6570\u636e\u7684\u79bb\u7ebf profile \u7ed3\u679c\u9009\u62e9\u8f83\u5c11\u4e13\u5bb6\u7684\u4e13\u5bb6\u9009\u62e9\u7b56\u7565<\/li>\n<\/ul>\n<p><em>\u6839\u636e KTransformers \u56e2\u961f\u5bf9 DeepSeekV2\u3001DeepSeekV3 \u548c DeepSeekR1 \u7684\u7814\u7a76\uff0c<br \/>\n\u5f53\u7a0d\u5fae\u51cf\u5c11\u63a8\u7406\u4e2d\u6fc0\u6d3b\u7684\u4e13\u5bb6\u6570\u91cf\u65f6\uff0c<br \/>\n\u8f93\u51fa\u8d28\u91cf\u4e0d\u4f1a\u6539\u53d8\u3002\u4f46\u662f\u89e3\u7801\u548c\u9884\u586b\u5145\u7684\u901f\u5ea6<br \/>\n\u4f1a\u52a0\u5feb\uff0c\u8fd9\u4ee4\u4eba\u9f13\u821e\u3002\u56e0\u6b64 KTransformers \u56e2\u961f\u7684\u6f14\u793a\u5229\u7528\u4e86\u8fd9\u4e00\u53d1\u73b0<\/em>\u00a0\u770b\u6765\u201c\u4e13\u5bb6\u9009\u62e9\u7b56\u7565\u201d\u662f\u63d0\u901f\u7684\u5173\u952e\uff0c\u4f46\u5982\u4f55\u4fdd\u8bc1\u8f93\u51fa\u8d28\u91cf\u4e0d\u4e0b\u964d\uff0c\u8fd8\u9700\u8981\u66f4\u591a\u6d4b\u8bd5\u548c\u9a8c\u8bc1\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>\u5982\u4f55\u8fd0\u884c<\/h2>\n<h3>V0.2 \u6f14\u793a<\/h3>\n<h4>\u5355\u8def\u7248\u672c (32 \u6838)<\/h4>\n<p>KTransformers \u56e2\u961f\u7684\u00a0<code>local_chat<\/code>\u00a0\u6d4b\u8bd5\u547d\u4ee4\u662f\uff1a<\/p>\n<pre><code>git clone https:\/\/github.com\/kvcache-ai\/ktransformers.git\r\ncd ktransformers\r\nnumactl -N 1 -m 1 python .\/ktransformers\/local_chat.py --model_path &lt;your model path&gt; --gguf_path &lt;your gguf path&gt;  --prompt_file &lt;your prompt txt file&gt;  --cpu_infer 33  --cache_lens 1536\r\n&lt;\u5f53\u770b\u5230 chat \u65f6\uff0c\u6309 Enter \u952e\u52a0\u8f7d\u6587\u672c prompt_file&gt;\r\n<\/code><\/pre>\n<p>&lt;your model path&gt; \u53ef\u4ee5\u662f\u672c\u5730\u8def\u5f84\uff0c\u4e5f\u53ef\u4ee5\u662f\u4ece\u5728\u7ebf hugging face \u8bbe\u7f6e\u7684\u8def\u5f84\uff0c\u4f8b\u5982 deepseek-ai\/DeepSeek-V3\u3002\u5982\u679c\u5728\u7ebf\u9047\u5230\u8fde\u63a5\u95ee\u9898\uff0c\u8bf7\u5c1d\u8bd5\u4f7f\u7528\u955c\u50cf (hf-mirror.com)<\/p>\n<p>&lt;your gguf path&gt; \u4e5f\u53ef\u4ee5\u662f\u5728\u7ebf\u8def\u5f84\uff0c\u4f46\u7531\u4e8e\u5b83\u5f88\u5927\uff0cKTransformers \u56e2\u961f\u5efa\u8bae\u60a8\u4e0b\u8f7d\u5b83\u5e76\u5c06\u6a21\u578b\u91cf\u5316\u4e3a\u60a8\u60f3\u8981\u7684\u683c\u5f0f<\/p>\n<p>\u547d\u4ee4\u00a0<code>numactl -N 1 -m 1<\/code>\u00a0\u65e8\u5728\u907f\u514d NUMA \u8282\u70b9\u4e4b\u95f4\u7684\u6570\u636e\u4f20\u8f93<\/p>\n<h4>\u53cc\u8def\u7248\u672c (64 \u6838)<\/h4>\n<p>\u5728\u5b89\u88c5\u4e4b\u524d (\u4f7f\u7528 install.sh \u6216\u00a0<code>make dev_install<\/code>)\uff0c\u901a\u8fc7\u00a0<code>export USE_NUMA=1<\/code>\u00a0\u8bbe\u7f6e\u73af\u5883\u53d8\u91cf\u00a0<code>USE_NUMA=1<\/code>\u00a0(\u5982\u679c\u5df2\u5b89\u88c5\uff0c\u8bf7\u5728\u8bbe\u7f6e\u6b64\u73af\u5883\u53d8\u91cf\u7684\u60c5\u51b5\u4e0b\u91cd\u65b0\u5b89\u88c5)<\/p>\n<p>KTransformers \u56e2\u961f\u7684\u00a0<code>local_chat<\/code>\u00a0\u6d4b\u8bd5\u547d\u4ee4\u662f\uff1a<\/p>\n<pre><code>git clone https:\/\/github.com\/kvcache-ai\/ktransformers.git\r\ncd ktransformers\r\nexport USE_NUMA=1\r\nmake dev_install # or sh .\/install.sh\r\npython .\/ktransformers\/local_chat.py --model_path &lt;your model path&gt; --gguf_path &lt;your gguf path&gt;  --prompt_file &lt;your prompt txt file&gt;  --cpu_infer 65  --cache_lens 1536\r\n&lt;\u5f53\u770b\u5230 chat \u65f6\uff0c\u6309 Enter \u952e\u52a0\u8f7d\u6587\u672c prompt_file&gt;\r\n<\/code><\/pre>\n<p>\u53c2\u6570\u7684\u542b\u4e49\u76f8\u540c\u3002\u4f46\u7531\u4e8e KTransformers \u56e2\u961f\u4f7f\u7528\u53cc\u8def\uff0c\u56e0\u6b64\u5c06\u00a0<code>cpu_infer<\/code>\u00a0\u8bbe\u7f6e\u4e3a 65<\/p>\n<h3>V0.3 \u6f14\u793a<\/h3>\n<h4>\u53cc\u8def\u7248\u672c (64 \u6838)<\/h4>\n<p>KTransformers \u56e2\u961f\u7684\u00a0<code>local_chat<\/code>\u00a0\u6d4b\u8bd5\u547d\u4ee4\u662f\uff1a<\/p>\n<pre><code>wget https:\/\/github.com\/kvcache-ai\/ktransformers\/releases\/download\/v0.1.4\/ktransformers-0.3.0rc0+cu126torch26fancy-cp311-cp311-linux_x86_64.whl\r\npip install .\/ktransformers-0.3.0rc0+cu126torch26fancy-cp311-cp311-linux_x86_64.whl\r\npython -m ktransformers.local_chat --model_path &lt;your model path&gt; --gguf_path &lt;your gguf path&gt;  --prompt_file &lt;your prompt txt file&gt;  --cpu_infer 65  --cache_lens 1536\r\n&lt;\u5f53\u770b\u5230 chat \u65f6\uff0c\u6309 Enter \u952e\u52a0\u8f7d\u6587\u672c prompt_file&gt;\r\n<\/code><\/pre>\n<p>\u53c2\u6570\u7684\u542b\u4e49\u4e0e V0.2 \u76f8\u540c\u3002\u4f46\u7531\u4e8e KTransformers \u56e2\u961f\u4f7f\u7528\u53cc\u8def\uff0c\u56e0\u6b64\u5c06\u00a0<code>cpu_infer<\/code>\u00a0\u8bbe\u7f6e\u4e3a 65<\/p>\n<p>&nbsp;<\/p>\n<h2>\u4e00\u4e9b\u89e3\u91ca<\/h2>\n<ol>\n<li>KTransformers \u56e2\u961f\u8fd8\u5e0c\u671b\u8fdb\u4e00\u6b65\u5229\u7528 Xeon Gold CPU \u4e0a\u7684\u4e24\u4e2a NUMA \u8282\u70b9\u3002<br \/>\n\u4e3a\u4e86\u907f\u514d\u8282\u70b9\u4e4b\u95f4\u6570\u636e\u4f20\u8f93\u7684\u6210\u672c\uff0cKTransformers \u56e2\u961f\u5728<br \/>\n\u4e24\u4e2a\u8282\u70b9\u4e0a\u201c\u590d\u5236\u201d\u5173\u952e\u77e9\u9635\uff0c\u8fd9\u4f1a\u6d88\u8017\u66f4\u591a\u5185\u5b58\uff0c\u4f46\u4f1a\u52a0\u901f\u9884\u586b\u5145\u548c\u89e3\u7801\u8fc7\u7a0b\u3002<br \/>\n\u4f46\u662f\u8fd9\u79cd\u65b9\u6cd5\u5360\u7528\u5927\u91cf\u5185\u5b58\uff0c\u5e76\u4e14\u5728\u52a0\u8f7d\u6743\u91cd\u65f6\u901f\u5ea6\u5f88\u6162\uff0c\u56e0\u6b64\u5728\u52a0\u8f7d\u65f6\u8bf7\u8010\u5fc3\u7b49\u5f85<br \/>\n\u5e76\u76d1\u63a7\u5185\u5b58\u4f7f\u7528\u60c5\u51b5\u3002KTransformers \u56e2\u961f\u5c06\u4f18\u5316\u8fd9\u79cd\u5de8\u5927\u7684\u5185\u5b58\u5f00\u9500\u3002\u656c\u8bf7\u5173\u6ce8~ \u8fd9\u79cd\u201c\u590d\u5236\u201d\u77e9\u9635\u7684\u65b9\u6cd5\u867d\u7136\u80fd\u63d0\u901f\uff0c\u4f46\u5185\u5b58\u5360\u7528\u786e\u5b9e\u662f\u4e2a\u95ee\u9898\uff0c\u671f\u5f85 KTransformers \u56e2\u961f\u540e\u7eed\u7684\u4f18\u5316\u65b9\u6848\u3002<\/li>\n<li>\u547d\u4ee4\u53c2\u6570\u00a0<code>--cpu_infer 65<\/code>\u00a0\u6307\u5b9a\u8981\u4f7f\u7528\u7684\u6838\u5fc3\u6570 (\u8d85\u8fc7\u7269\u7406\u6838\u5fc3\u6570\u4e5f\u53ef\u4ee5\uff0c<br \/>\n\u4f46\u5e76\u975e\u8d8a\u591a\u8d8a\u597d\u3002\u5c06\u5176\u8c03\u6574\u5230\u7565\u4f4e\u4e8e\u5b9e\u9645\u6838\u5fc3\u6570\u5373\u53ef)<\/li>\n<li>\u4e3a\u4ec0\u4e48\u91c7\u7528 CPU\/GPU \u6df7\u5408\u63a8\u7406\uff1f<br \/>\n<a href=\"https:\/\/www.kdjingpai.com\/deepseek-chatshena\/\">DeepSeek<\/a> \u7684 MLA \u7b97\u5b50\u662f\u8ba1\u7b97\u5bc6\u96c6\u578b\u7684\u3002\u867d\u7136\u5b8c\u5168\u5728 CPU \u4e0a\u8fd0\u884c\u662f\u53ef\u80fd\u7684\uff0c\u4f46\u5c06\u7e41\u91cd\u7684\u8ba1\u7b97\u5378\u8f7d\u5230 GPU \u53ef\u4ee5\u5927\u5e45\u63d0\u5347\u6027\u80fd\u3002CPU \u8d1f\u8d23\u4e13\u5bb6\u8ba1\u7b97\uff0cGPU \u8d1f\u8d23 MLA\/KVCache\uff0c\u8fd9\u79cd\u6df7\u5408\u63a8\u7406\u7684\u7b56\u7565\u770b\u8d77\u6765\u5f88\u806a\u660e\uff0c\u5145\u5206\u5229\u7528\u4e86 CPU \u548c GPU \u7684\u4f18\u52bf\u3002<\/li>\n<li>\u901f\u5ea6\u63d0\u5347\u6765\u81ea\u54ea\u91cc\uff1f\n<ul>\n<li>\u4e13\u5bb6\u5378\u8f7d (Expert Offload): \u4e0e\u4f20\u7edf\u7684\u57fa\u4e8e\u5c42\u6216 KVCache \u7684\u5378\u8f7d (\u5982 llama.cpp \u4e2d\u6240\u89c1) \u4e0d\u540c\uff0cKTransformers \u56e2\u961f\u5c06\u4e13\u5bb6\u8ba1\u7b97\u5378\u8f7d\u5230 CPU\uff0c\u5c06 MLA\/KVCache \u5378\u8f7d\u5230 GPU\uff0c\u4e0e DeepSeek \u7684\u67b6\u6784\u5b8c\u7f8e\u5951\u5408\uff0c\u4ee5\u5b9e\u73b0\u6700\u4f73\u6548\u7387\u3002<\/li>\n<li>Intel AMX \u4f18\u5316 \u2013 KTransformers \u56e2\u961f\u57fa\u4e8e AMX \u52a0\u901f\u7684\u5185\u6838\u7ecf\u8fc7\u7cbe\u5fc3\u8c03\u4f18\uff0c\u8fd0\u884c\u901f\u5ea6\u6bd4\u73b0\u6709\u7684 llama.cpp \u5b9e\u73b0\u5feb\u6570\u500d\u3002KTransformers \u56e2\u961f\u8ba1\u5212\u5728\u6e05\u7406\u540e\u5f00\u6e90\u6b64\u5185\u6838\uff0c\u5e76\u6b63\u5728\u8003\u8651\u5411\u4e0a\u6e38 llama.cpp \u8d21\u732e\u4ee3\u7801\u3002AMX \u6307\u4ee4\u96c6\u7684\u52a0\u6301\uff0c\u770b\u6765\u662f KTransformers \u63d0\u901f\u7684\u5173\u952e\u56e0\u7d20\u4e4b\u4e00\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\u4e3a\u4ec0\u4e48\u9009\u62e9 Intel CPU\uff1f<br \/>\nIntel \u662f\u76ee\u524d\u552f\u4e00\u652f\u6301\u7c7b\u4f3c AMX \u6307\u4ee4\u7684 CPU \u4f9b\u5e94\u5546\uff0c\u4e0e\u4ec5\u652f\u6301 AVX \u7684\u66ff\u4ee3\u65b9\u6848\u76f8\u6bd4\uff0cAMX \u6307\u4ee4\u53ef\u63d0\u4f9b\u660e\u663e\u66f4\u597d\u7684\u6027\u80fd\u3002\u770b\u6765\u76ee\u524d\u60f3\u4f53\u9a8c KTransformers \u7684\u6700\u4f73\u6027\u80fd\uff0cIntel CPU \u662f\u4e0d\u4e8c\u4e4b\u9009\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u6458\u8981 2025 \u5e74 2 \u6708 10 \u65e5: \u5728\u5355 GPU (24GB \u663e\u5b58) \/ \u591a GPU \u548c 382GB \u5185\u5b58\u4e0a\u652f\u6301 DeepseekR1 \u548c V3\uff0c\u901f\u5ea6\u63d0\u5347\u9ad8\u8fbe 3~28 \u500d\u3002 \u5927\u5bb6\u597d\uff0cKTransformers \u56e2\u961f (\u524d\u8eab\u4e3a CPU\/&#8230;<\/p>\n","protected":false},"author":1,"featured_media":61808,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20,46],"tags":[],"class_list":["post-20794","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","category-news"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/20794","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/comments?post=20794"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/20794\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media\/61808"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media?parent=20794"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/categories?post=20794"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/tags?post=20794"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}