{"id":9420,"date":"2025-06-10T19:49:50","date_gmt":"2025-06-10T11:49:50","guid":{"rendered":"\/?p=9420"},"modified":"2025-06-10T19:49:50","modified_gmt":"2025-06-10T11:49:50","slug":"%e5%bd%93%e5%89%8d%e4%b8%bb%e6%b5%81%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%96%87%e4%bb%b6%e6%a0%bc%e5%bc%8f%e8%a7%a3%e6%9e%90%e5%af%b9%e6%af%94","status":"publish","type":"post","link":"\/?p=9420","title":{"rendered":"\u5f53\u524d\u4e3b\u6d41\u5927\u6a21\u578b\u6587\u4ef6\u683c\u5f0f\u89e3\u6790\u5bf9\u6bd4"},"content":{"rendered":"<h3>1. HuggingFace Transformers \u683c\u5f0f\uff08<code>.bin<\/code> \/ <code>.safetensors<\/code>\uff09<\/h3>\n<table>\n<thead>\n<tr>\n<th>\u9879\u76ee<\/th>\n<th>\u8bf4\u660e<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>\u7528\u9014<\/strong><\/td>\n<td>\u8bad\u7ec3\/\u63a8\u7406\u901a\u7528\u683c\u5f0f<\/td>\n<\/tr>\n<tr>\n<td><strong>\u652f\u6301\u5e73\u53f0<\/strong><\/td>\n<td>HuggingFace Transformers, vLLM, Text Generation Inference (TGI), DeepSpeed \u7b49<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>\u4f18\u70b9\uff1a<\/strong><br \/>\n\u2705 \u6a21\u578b\u751f\u6001\u6700\u4e30\u5bcc\uff08LLaMA\u3001Baichuan\u3001ChatGLM \u7b49\uff09<br \/>\n\u2705 \u652f\u6301\u8bad\u7ec3+\u63a8\u7406<br \/>\n\u2705 \u548c HuggingFace Hub \u517c\u5bb9\u597d<br \/>\n\u2705 <code>.safetensors<\/code> \u66f4\u5b89\u5168\uff0c\u652f\u6301 mmap \u63d0\u5347\u52a0\u8f7d\u901f\u5ea6  <\/p>\n<p><strong>\u7f3a\u70b9\uff1a<\/strong><br \/>\n\u274c <code>.bin<\/code> \u683c\u5f0f\u53ef\u80fd\u5b58\u5728\u5b89\u5168\u98ce\u9669\uff08Pickle \u6267\u884c\uff09<br \/>\n\u274c \u6a21\u578b\u672a\u538b\u7f29\uff0c\u5360\u7528\u663e\u5b58\u5927<br \/>\n\u274c \u542f\u52a8\u52a0\u8f7d\u901f\u5ea6\u8f83\u6162\uff0c\u5c24\u5176\u662f\u5927\u578b\u6a21\u578b  <\/p>\n<p><strong>\u63a8\u8350\u7528\u9014\uff1a<\/strong><br \/>\n\u4e3b\u6d41\u6a21\u578b\u8bad\u7ec3\u3001vLLM\/TGI\u9ad8\u6027\u80fd\u63a8\u7406\u90e8\u7f72\u3001\u7814\u7a76\u590d\u73b0\u7b49<\/p>\n<hr \/>\n<h3>\ud83e\udde9 2. GGUF \u683c\u5f0f\uff08<code>.gguf<\/code>\uff0cGGML \u7edf\u4e00\u683c\u5f0f\uff09<\/h3>\n<table>\n<thead>\n<tr>\n<th>\u9879\u76ee<\/th>\n<th>\u8bf4\u660e<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>\u7528\u9014<\/strong><\/td>\n<td>\u672c\u5730\u4f4e\u8d44\u6e90\u73af\u5883\u7684\u91cf\u5316\u63a8\u7406<\/td>\n<\/tr>\n<tr>\n<td><strong>\u652f\u6301\u5e73\u53f0<\/strong><\/td>\n<td>llama.cpp\u3001koboldcpp\u3001LM Studio\u3001Ollama\u3001text-generation-webui\u3001MLC-LLM<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>\u4f18\u70b9\uff1a<\/strong><br \/>\n\u2705 \u652f\u6301 8\/6\/5\/4\/3bit \u91cf\u5316\uff0c\u663e\u8457\u51cf\u5c0f\u6a21\u578b\u4f53\u79ef<br \/>\n\u2705 \u53ef\u5728 CPU\u3001\u672c\u5730 GPU\u3001\u751a\u81f3\u5b89\u5353\/iOS \u4e0a\u8fd0\u884c<br \/>\n\u2705 \u542f\u52a8\u5feb\uff0c\u5360\u7528\u5c0f\uff0c\u9002\u5408\u79bb\u7ebf\/\u79fb\u52a8\u7aef\u90e8\u7f72<br \/>\n\u2705 \u4e0e llama.cpp\u3001Ollama \u5b8c\u7f8e\u517c\u5bb9  <\/p>\n<p><strong>\u7f3a\u70b9\uff1a<\/strong><br \/>\n\u274c \u4e0d\u652f\u6301\u8bad\u7ec3<br \/>\n\u274c \u4e0d\u652f\u6301\u63a8\u7406\u5fae\u8c03\u540e\u7684\u6743\u91cd\uff08\u90e8\u5206\u91cf\u5316\u635f\u5931\u4fe1\u606f\uff09<br \/>\n\u274c \u6a21\u578b\u7ed3\u6784\u8f83\u56fa\u5b9a\uff0c\u529f\u80fd\u4e0d\u5982 Transformers \u4e30\u5bcc  <\/p>\n<p><strong>\u63a8\u8350\u7528\u9014\uff1a<\/strong><br \/>\n\u8f7b\u91cf\u672c\u5730\u63a8\u7406\u3001\u65e0 GPU \u73af\u5883\u3001\u79fb\u52a8\u7aef\u3001\u4fbf\u643a\u5f0f AI \u52a9\u624b\u7b49\u573a\u666f<\/p>\n<hr \/>\n<h3>\ud83e\udde9 3. PyTorch \u539f\u751f\u683c\u5f0f\uff08<code>.pt<\/code>, <code>.pth<\/code>\uff09<\/h3>\n<table>\n<thead>\n<tr>\n<th>\u9879\u76ee<\/th>\n<th>\u8bf4\u660e<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>\u7528\u9014<\/strong><\/td>\n<td>\u8bad\u7ec3\u4e0e\u5b9e\u9a8c\u7814\u7a76<\/td>\n<\/tr>\n<tr>\n<td><strong>\u652f\u6301\u5e73\u53f0<\/strong><\/td>\n<td>PyTorch \u539f\u751f\u3001Fairseq\u3001OpenNMT\u3001DeepSpeed \u7b49<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>\u4f18\u70b9\uff1a<\/strong><br \/>\n\u2705 \u539f\u751f\u4fdd\u5b58 PyTorch \u6a21\u578b\u6743\u91cd\u3001\u4f18\u5316\u5668\u3001\u8bad\u7ec3\u72b6\u6001<br \/>\n\u2705 \u7075\u6d3b\u6027\u9ad8\uff0c\u9002\u5408\u7814\u7a76\u548c\u81ea\u5b9a\u4e49\u6a21\u578b\u7ed3\u6784<br \/>\n\u2705 \u4e0e PyTorch \u8bad\u7ec3\/\u5fae\u8c03\u6d41\u7a0b\u65e0\u7f1d\u96c6\u6210  <\/p>\n<p><strong>\u7f3a\u70b9\uff1a<\/strong><br \/>\n\u274c \u4e0d\u9002\u5408\u76f4\u63a5\u90e8\u7f72\u670d\u52a1\uff08\u7f3a\u4e4f\u6807\u51c6\u63a5\u53e3\uff09<br \/>\n\u274c \u52a0\u8f7d\u6162\uff0c\u5360\u7528\u9ad8<br \/>\n\u274c \u5b89\u5168\u6027\u5f31\uff08\u4f7f\u7528 pickle\uff09  <\/p>\n<p><strong>\u63a8\u8350\u7528\u9014\uff1a<\/strong><br \/>\n\u8bad\u7ec3\u9636\u6bb5\u6a21\u578b\u4fdd\u5b58\u3001\u4e2d\u95f4\u8c03\u8bd5\u3001\u81ea\u4e3b\u7814\u53d1\u6a21\u578b\u8bad\u7ec3\u6d41\u7a0b<\/p>\n<hr \/>\n<h3>\ud83e\udde9 4. Safetensors \u683c\u5f0f\uff08<code>.safetensors<\/code>\uff09<\/h3>\n<table>\n<thead>\n<tr>\n<th>\u9879\u76ee<\/th>\n<th>\u8bf4\u660e<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>\u7528\u9014<\/strong><\/td>\n<td>\u66ff\u4ee3 <code>.bin<\/code>\uff0c\u63d0\u9ad8\u5b89\u5168\u6027\u4e0e\u52a0\u8f7d\u6548\u7387<\/td>\n<\/tr>\n<tr>\n<td><strong>\u652f\u6301\u5e73\u53f0<\/strong><\/td>\n<td>HuggingFace Transformers\u3001vLLM\u3001text-generation-webui \u7b49<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>\u4f18\u70b9\uff1a<\/strong><br \/>\n\u2705 \u96f6\u4fe1\u4efb\u5b89\u5168\u683c\u5f0f\uff0c\u9632\u6b62 pickle \u6267\u884c<br \/>\n\u2705 \u52a0\u8f7d\u66f4\u5feb\uff0c\u652f\u6301\u5185\u5b58\u6620\u5c04\uff08mmap\uff09<br \/>\n\u2705 \u652f\u6301\u6a21\u578b\u5207\u7247\u5e76\u884c\u52a0\u8f7d  <\/p>\n<p><strong>\u7f3a\u70b9\uff1a<\/strong><br \/>\n\u274c \u4e0d\u652f\u6301\u8bad\u7ec3\u4e2d\u95f4\u72b6\u6001\uff08\u5982\u4f18\u5316\u5668\u53c2\u6570\uff09<br \/>\n\u274c \u4e00\u4e9b\u8001\u5de5\u5177\u6216\u811a\u672c\u5c1a\u4e0d\u517c\u5bb9  <\/p>\n<p><strong>\u63a8\u8350\u7528\u9014\uff1a<\/strong><br \/>\n\u5b89\u5168\u90e8\u7f72\u3001\u4e91\u670d\u52a1\u63a8\u7406\u3001\u9ad8\u6027\u80fd\u6a21\u578b\u52a0\u8f7d\uff08\u63a8\u8350\u66ff\u4ee3 <code>.bin<\/code> \u4f7f\u7528\uff09<\/p>\n<hr \/>\n<h3>\ud83e\udde9 5. ONNX \u683c\u5f0f\uff08<code>.onnx<\/code>\uff09<\/h3>\n<table>\n<thead>\n<tr>\n<th>\u9879\u76ee<\/th>\n<th>\u8bf4\u660e<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>\u7528\u9014<\/strong><\/td>\n<td>\u8de8\u5e73\u53f0\u3001\u63a8\u7406\u4f18\u5316\u90e8\u7f72<\/td>\n<\/tr>\n<tr>\n<td><strong>\u652f\u6301\u5e73\u53f0<\/strong><\/td>\n<td>ONNX Runtime, TensorRT, OpenVINO, DeepSparse \u7b49<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>\u4f18\u70b9\uff1a<\/strong><br \/>\n\u2705 \u8de8\u5e73\u53f0\u90e8\u7f72\uff1ax86\u3001ARM\u3001Web\u3001\u8fb9\u7f18\u8bbe\u5907<br \/>\n\u2705 \u652f\u6301\u591a\u79cd\u786c\u4ef6\u63a8\u7406\u5f15\u64ce<br \/>\n\u2705 \u63a8\u7406\u901f\u5ea6\u5feb\uff0c\u9002\u5408\u5c0f\u578b\u6a21\u578b  <\/p>\n<p><strong>\u7f3a\u70b9\uff1a<\/strong><br \/>\n\u274c LLM \u8f6c\u6362\u590d\u6742\uff0c\u517c\u5bb9\u6027\u5dee\uff08\u5982 LLaMA\u3001ChatGLM \u8f6c\u6362\u5bb9\u6613\u5931\u8d25\uff09<br \/>\n\u274c \u52a8\u6001\u6a21\u578b\u652f\u6301\u8f83\u5f31<br \/>\n\u274c \u4ec5\u9002\u5408\u63a8\u7406\uff0c\u96be\u4ee5\u5fae\u8c03  <\/p>\n<p><strong>\u63a8\u8350\u7528\u9014\uff1a<\/strong><br \/>\n\u5c0f\u6a21\u578b\u90e8\u7f72\u5230\u8fb9\u7f18\/\u6d4f\u89c8\u5668\/\u5d4c\u5165\u5f0f\u8bbe\u5907\uff0c\u5982 BERT\u3001TinyGPT\u3001Whisper \u7b49\u573a\u666f<\/p>\n<hr \/>\n<h3>\ud83e\udde9 6. TensorFlow \u7cfb\u5217\uff08<code>.pb<\/code>, <code>.ckpt<\/code>, SavedModel\uff09<\/h3>\n<table>\n<thead>\n<tr>\n<th>\u9879\u76ee<\/th>\n<th>\u8bf4\u660e<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>\u7528\u9014<\/strong><\/td>\n<td>TensorFlow \u6a21\u578b\u8bad\u7ec3\u4e0e\u90e8\u7f72<\/td>\n<\/tr>\n<tr>\n<td><strong>\u652f\u6301\u5e73\u53f0<\/strong><\/td>\n<td>TensorFlow\u3001TF Lite\u3001Google Cloud AI<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>\u4f18\u70b9\uff1a<\/strong><br \/>\n\u2705 \u9002\u7528\u4e8e Google \u751f\u6001\uff08\u5982 GCP AI \u5e73\u53f0\uff09<br \/>\n\u2705 TF Lite \u53ef\u90e8\u7f72\u5230\u79fb\u52a8\u7aef<br \/>\n\u2705 SavedModel \u683c\u5f0f\u7ed3\u6784\u6e05\u6670  <\/p>\n<p><strong>\u7f3a\u70b9\uff1a<\/strong><br \/>\n\u274c \u5f53\u524d\u5927\u6a21\u578b\u4e3b\u6d41\u751f\u6001\u5df2\u8f6c\u5411 PyTorch<br \/>\n\u274c \u5de5\u5177\u94fe\u7e41\u6742\uff0c\u683c\u5f0f\u7e41\u591a<br \/>\n\u274c HuggingFace \u8f6c\u6362\u652f\u6301\u4e0d\u5b8c\u5584  <\/p>\n<p><strong>\u63a8\u8350\u7528\u9014\uff1a<\/strong><br \/>\n\u5728 TensorFlow \u9879\u76ee\u4e2d\u90e8\u7f72 BERT\u3001T5\u3001Tiny GPT \u7b49\u6a21\u578b\uff0c\u9002\u7528\u4e8e\u79fb\u52a8\u7aef\u5e94\u7528\u6216\u8c37\u6b4c\u5e73\u53f0\u7528\u6237<\/p>\n<hr \/>\n<h3>\ud83d\udcca \u6c47\u603b\u5bf9\u6bd4\u8868<\/h3>\n<table>\n<thead>\n<tr>\n<th>\u683c\u5f0f<\/th>\n<th>\u652f\u6301\u63a8\u7406<\/th>\n<th>\u652f\u6301\u8bad\u7ec3<\/th>\n<th>\u91cf\u5316\u652f\u6301<\/th>\n<th>\u5b89\u5168\u6027<\/th>\n<th>\u90e8\u7f72\u6613\u7528\u6027<\/th>\n<th>\u63a8\u8350\u573a\u666f<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><code>.bin<\/code><\/td>\n<td>\u2705<\/td>\n<td>\u2705<\/td>\n<td>\u274c<\/td>\n<td>\u26a0\ufe0f \u4f4e\uff08pickle\uff09<\/td>\n<td>\u2b50\u2b50\u2b50\u2b50<\/td>\n<td>\u901a\u7528\u3001\u5fae\u8c03\u3001vLLM<\/td>\n<\/tr>\n<tr>\n<td><code>.safetensors<\/code><\/td>\n<td>\u2705<\/td>\n<td>\u274c<\/td>\n<td>\u26a0\ufe0f \u90e8\u5206\u652f\u6301<\/td>\n<td>\u2705 \u9ad8<\/td>\n<td>\u2b50\u2b50\u2b50\u2b50\u2b50<\/td>\n<td>\u9ad8\u6027\u80fd\u5b89\u5168\u90e8\u7f72<\/td>\n<\/tr>\n<tr>\n<td><code>.gguf<\/code><\/td>\n<td>\u2705<\/td>\n<td>\u274c<\/td>\n<td>\u2705 \u5f3a<\/td>\n<td>\u2705 \u9ad8<\/td>\n<td>\u2b50\u2b50\u2b50\u2b50\u2b50<\/td>\n<td>\u672c\u5730\u90e8\u7f72\uff0c\u4f4e\u8d44\u6e90\u8fd0\u884c<\/td>\n<\/tr>\n<tr>\n<td><code>.pt\/.pth<\/code><\/td>\n<td>\u26a0\ufe0f \u90e8\u5206<\/td>\n<td>\u2705<\/td>\n<td>\u274c<\/td>\n<td>\u26a0\ufe0f \u4f4e\uff08pickle\uff09<\/td>\n<td>\u2b50\u2b50<\/td>\n<td>\u81ea\u7814\u8bad\u7ec3\u6d41\u7a0b<\/td>\n<\/tr>\n<tr>\n<td><code>.onnx<\/code><\/td>\n<td>\u2705<\/td>\n<td>\u274c<\/td>\n<td>\u26a0\ufe0f \u6709\u9650\u652f\u6301<\/td>\n<td>\u2705 \u9ad8<\/td>\n<td>\u2b50\u2b50\u2b50\u2b50<\/td>\n<td>\u63a8\u7406\u4f18\u5316\u3001Web\/\u79fb\u52a8\u7aef<\/td>\n<\/tr>\n<tr>\n<td><code>.pb\/.ckpt<\/code><\/td>\n<td>\u2705<\/td>\n<td>\u2705<\/td>\n<td>\u26a0\ufe0f \u5f31<\/td>\n<td>\u2705 \u4e2d<\/td>\n<td>\u2b50\u2b50<\/td>\n<td>TF \u9879\u76ee\uff0c\u8c37\u6b4c\u5e73\u53f0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n","protected":false},"excerpt":{"rendered":"<p>1. HuggingFace Transformers \u683c\u5f0f\uff08.bin \/ .safetensors\uff09 \u9879\u76ee \u8bf4\u660e \u7528\u9014 \u8bad\u7ec3\/\u63a8\u7406\u901a\u7528   \u2026 &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ngg_post_thumbnail":0},"categories":[224],"tags":[],"_links":{"self":[{"href":"\/index.php?rest_route=\/wp\/v2\/posts\/9420"}],"collection":[{"href":"\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=9420"}],"version-history":[{"count":1,"href":"\/index.php?rest_route=\/wp\/v2\/posts\/9420\/revisions"}],"predecessor-version":[{"id":9421,"href":"\/index.php?rest_route=\/wp\/v2\/posts\/9420\/revisions\/9421"}],"wp:attachment":[{"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=9420"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=9420"},{"taxonomy":"post_tag","embeddable":true,"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=9420"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}