{"id":29343,"date":"2025-01-29T16:02:52","date_gmt":"2025-01-29T08:02:52","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=29343"},"modified":"2025-03-25T16:06:37","modified_gmt":"2025-03-25T08:06:37","slug":"qwen25-vl","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/de\/qwen25-vl\/","title":{"rendered":"Qwen2.5-VL\uff1a\u652f\u6301\u56fe\u50cf\u89c6\u9891\u6587\u6863\u89e3\u6790\u7684\u5f00\u6e90\u591a\u6a21\u6001\u5927\u6a21\u578b"},"content":{"rendered":"<p>Qwen2.5-VL \u662f\u963f\u91cc\u5df4\u5df4\u4e91\uff08Alibaba Cloud\uff09Qwen \u56e2\u961f\u5f00\u53d1\u7684\u5f00\u6e90\u591a\u6a21\u6001\u5927\u6a21\u578b\u3002\u5b83\u80fd\u540c\u65f6\u5904\u7406\u6587\u672c\u3001\u56fe\u50cf\u3001\u89c6\u9891\u548c\u6587\u6863\uff0c\u662f Qwen2-VL \u7684\u5347\u7ea7\u7248\uff0c\u57fa\u4e8e Qwen2.5 \u8bed\u8a00\u6a21\u578b\u6784\u5efa\u3002\u5b98\u65b9\u79f0\uff0c\u5b83\u5728\u6587\u6863\u89e3\u6790\u3001\u89c6\u9891\u7406\u89e3\u548c\u667a\u80fd\u4ee3\u7406\u529f\u80fd\u4e0a\u5927\u5e45\u63d0\u5347\uff0c\u652f\u6301 3B\u30017B\u300132B \u548c 72B \u56db\u79cd\u53c2\u6570\u89c4\u6a21\uff0c\u6ee1\u8db3\u4ece\u4e2a\u4eba\u7535\u8111\u5230\u670d\u52a1\u5668\u7684\u591a\u79cd\u9700\u6c42\u3002\u9879\u76ee\u6258\u7ba1\u5728 GitHub\uff0c\u91c7\u7528 Apache 2.0 \u8bb8\u53ef\u8bc1\uff0c\u514d\u8d39\u5f00\u653e\u6e90\u7801\u3002Qwen2.5-VL \u5728\u591a\u4e2a\u6d4b\u8bd5\u4e2d\u8868\u73b0\u4f18\u5f02\uff0c\u90e8\u5206\u6307\u6807\u8d85\u8fc7\u95ed\u6e90\u6a21\u578b\uff0c\u9002\u5408\u5f00\u53d1\u8005\u6253\u9020\u667a\u80fd\u5de5\u5177\uff0c\u5982\u6587\u6863\u63d0\u53d6\u5668\u3001\u89c6\u9891\u5206\u6790\u5668\u6216\u8bbe\u5907\u64cd\u4f5c\u52a9\u624b\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter  wp-image-29345\" title=\"Qwen2.5-VL\uff1a\u652f\u6301\u56fe\u50cf\u89c6\u9891\u6587\u6863\u89e3\u6790\u7684\u5f00\u6e90\u591a\u6a21\u6001\u5927\u6a21\u578b-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/d3c12215cb28063.png\" alt=\"Qwen2.5-VL\uff1a\u652f\u6301\u56fe\u50cf\u89c6\u9891\u6587\u6863\u89e3\u6790\u7684\u5f00\u6e90\u591a\u6a21\u6001\u5927\u6a21\u578b-1\" width=\"820\" height=\"541\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/d3c12215cb28063.png 956w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/d3c12215cb28063-768x507.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/d3c12215cb28063-18x12.png 18w\" sizes=\"auto, (max-width: 820px) 100vw, 820px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<ul>\n<li>\u8bc6\u522b\u56fe\u50cf\u4e2d\u7684\u7269\u4f53\u3001\u6587\u5b57\u548c\u5e03\u5c40\uff0c\u652f\u6301\u81ea\u7136\u573a\u666f\u548c\u591a\u8bed\u8a00\u3002<\/li>\n<li>\u7406\u89e3\u8d85\u957f\u89c6\u9891\uff08\u8d85\u8fc7\u4e00\u5c0f\u65f6\uff09\uff0c\u5b9a\u4f4d\u79d2\u7ea7\u4e8b\u4ef6\u7247\u6bb5\u3002<\/li>\n<li>\u89e3\u6790\u590d\u6742\u6587\u6863\uff0c\u63d0\u53d6\u624b\u5199\u6587\u5b57\u3001\u8868\u683c\u3001\u56fe\u8868\u548c\u5316\u5b66\u516c\u5f0f\u3002<\/li>\n<li>\u901a\u8fc7\u89c6\u89c9\u548c\u6587\u5b57\u6307\u4ee4\u63a7\u5236\u7535\u8111\u6216\u624b\u673a\uff0c\u6267\u884c\u667a\u80fd\u4ee3\u7406\u4efb\u52a1\u3002<\/li>\n<li>\u8f93\u51fa\u7ed3\u6784\u5316\u6570\u636e\uff0c\u5982 JSON \u683c\u5f0f\u7684\u5750\u6807\u6216\u5c5e\u6027\u4fe1\u606f\u3002<\/li>\n<li>\u652f\u6301\u52a8\u6001\u5206\u8fa8\u7387\u548c\u5e27\u7387\u8c03\u6574\uff0c\u4f18\u5316\u89c6\u9891\u5904\u7406\u6548\u7387\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<h3>\u5b89\u88c5\u6d41\u7a0b<\/h3>\n<p>\u8981\u5728\u672c\u5730\u8fd0\u884c Qwen2.5-VL\uff0c\u9700\u51c6\u5907\u8f6f\u4ef6\u73af\u5883\u3002\u4ee5\u4e0b\u662f\u8be6\u7ec6\u6b65\u9aa4\uff1a<\/p>\n<ol>\n<li><strong>\u68c0\u67e5\u57fa\u7840\u73af\u5883<\/strong><br \/>\n\u9700\u8981 Python 3.8+ \u548c Git\u3002\u7ec8\u7aef\u8f93\u5165\u4ee5\u4e0b\u547d\u4ee4\u68c0\u67e5\uff1a<\/li>\n<\/ol>\n<pre><code>python --version\r\ngit --version\r\n<\/code><\/pre>\n<p>\u672a\u5b89\u88c5\u53ef\u4ece\u00a0Python \u5b98\u7f51\u00a0\u548c\u00a0Git \u5b98\u7f51\u00a0\u4e0b\u8f7d\u3002<\/p>\n<ol start=\"2\">\n<li><strong>\u4e0b\u8f7d\u4ee3\u7801<\/strong><br \/>\n\u5728\u7ec8\u7aef\u8fd0\u884c\u4ee5\u4e0b\u547d\u4ee4\u514b\u9686 GitHub \u4ed3\u5e93\uff1a<\/li>\n<\/ol>\n<pre><code>git clone https:\/\/github.com\/QwenLM\/Qwen2.5-VL.git\r\ncd Qwen2.5-VL\r\n<\/code><\/pre>\n<ol start=\"3\">\n<li><strong>\u5b89\u88c5\u4f9d\u8d56<\/strong><br \/>\n\u9879\u76ee\u9700\u8981\u7279\u5b9a Python \u5e93\uff0c\u8fd0\u884c\u4ee5\u4e0b\u547d\u4ee4\u5b89\u88c5\uff1a<\/li>\n<\/ol>\n<pre><code>pip install git+https:\/\/github.com\/huggingface\/transformers@f3f6c86582611976e72be054675e2bf0abb5f775\r\npip install accelerate\r\npip install qwen-vl-utils[decord]\r\npip install 'vllm&gt;0.7.2'\r\n<\/code><\/pre>\n<p>\u5982\u679c\u6709 GPU\uff0c\u5b89\u88c5\u652f\u6301 CUDA \u7684 PyTorch\uff1a<\/p>\n<pre><code>pip install torch torchvision torchaudio --index-url https:\/\/download.pytorch.org\/whl\/cu124\r\n<\/code><\/pre>\n<p>\u6ce8\uff1a<code>decord<\/code>\u00a0\u7528\u4e8e\u52a0\u901f\u89c6\u9891\u52a0\u8f7d\uff0c\u975e Linux \u7528\u6237\u53ef\u4ece\u00a0<a href=\"https:\/\/github.com\/dmlc\/decord\">decord GitHub<\/a>\u00a0\u6e90\u7801\u5b89\u88c5\u3002<\/p>\n<ol start=\"4\">\n<li><strong>\u542f\u52a8\u6a21\u578b<\/strong><br \/>\n\u4e0b\u8f7d\u5e76\u8fd0\u884c\u6a21\u578b\uff0c\u4f8b\u5982 7B \u7248\u672c\uff1a<\/li>\n<\/ol>\n<pre><code><a href=\"https:\/\/www.kdjingpai.com\/de\/vllm\/\">vllm<\/a> serve Qwen\/Qwen2.5-VL-7B-Instruct --port 8000 --host 0.0.0.0 --dtype bfloat16\r\n<\/code><\/pre>\n<p>\u6a21\u578b\u4f1a\u81ea\u52a8\u4ece Hugging Face \u4e0b\u8f7d\u5e76\u542f\u52a8\u672c\u5730\u670d\u52a1\u3002<\/p>\n<h3>\u5982\u4f55\u4f7f\u7528\u4e3b\u8981\u529f\u80fd<\/h3>\n<p>\u5b89\u88c5\u540e\uff0c\u53ef\u901a\u8fc7\u4ee3\u7801\u6216 Web \u754c\u9762\u64cd\u4f5c Qwen2.5-VL\u3002<\/p>\n<h4>\u56fe\u50cf\u8bc6\u522b<\/h4>\n<p>\u60f3\u8ba9\u6a21\u578b\u63cf\u8ff0\u56fe\u7247\uff1f\u65b0\u5efa\u00a0<code>image_test.py<\/code>\uff0c\u8f93\u5165\u4ee5\u4e0b\u4ee3\u7801\uff1a<\/p>\n<pre><code>from transformers import Qwen2VLForConditionalGeneration, AutoProcessor\r\nfrom qwen_vl_utils import process_vision_info\r\nfrom PIL import Image\r\nmodel = Qwen2VLForConditionalGeneration.from_pretrained(\"Qwen\/Qwen2.5-VL-7B-Instruct\", torch_dtype=\"auto\", device_map=\"auto\")\r\nprocessor = AutoProcessor.from_pretrained(\"Qwen\/Qwen2.5-VL-7B-Instruct\")\r\nmessages = [{\"role\": \"user\", \"content\": [{\"type\": \"image\", \"image\": \"\u56fe\u7247\u8def\u5f84\u6216URL\"}, {\"type\": \"text\", \"text\": \"\u63cf\u8ff0\u8fd9\u5f20\u56fe\u7247\"}]}]\r\ntext = processor.apply_chat_template(messages, tokenize=False)\r\ninputs = processor(text=[text], images=[Image.open(\"\u56fe\u7247\u8def\u5f84\u6216URL\")], padding=True, return_tensors=\"pt\").to(\"cuda\")\r\ngenerated_ids = model.generate(**inputs, max_new_tokens=128)\r\noutput = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]\r\nprint(output)\r\n<\/code><\/pre>\n<p>\u8fd0\u884c\u540e\uff0c\u6a21\u578b\u4f1a\u8f93\u51fa\u63cf\u8ff0\uff0c\u4f8b\u5982\uff1a\u201c\u56fe\u7247\u91cc\u6709\u4e00\u53ea\u732b\u5750\u5728\u7a97\u53f0\u4e0a\u3002\u201d<\/p>\n<h4>\u89c6\u9891\u7406\u89e3<\/h4>\n<p>\u5206\u6790\u89c6\u9891\u9700\u51c6\u5907\u672c\u5730\u6587\u4ef6\uff08\u5982\u00a0<code>video.mp4<\/code>\uff09\u3002\u7528\u4ee5\u4e0b\u4ee3\u7801\uff1a<\/p>\n<pre><code>messages = [{\"role\": \"user\", \"content\": [{\"type\": \"video\", \"video\": \"video.mp4\"}, {\"type\": \"text\", \"text\": \"\u603b\u7ed3\u89c6\u9891\u5185\u5bb9\"}]}]\r\ntext = processor.apply_chat_template(messages, tokenize=False)\r\ninputs = processor(text=[text], videos=[processor.process_video(\"video.mp4\")], padding=True, return_tensors=\"pt\").to(\"cuda\")\r\ngenerated_ids = model.generate(**inputs, max_new_tokens=128)\r\noutput = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]\r\nprint(output)\r\n<\/code><\/pre>\n<p>\u8f93\u51fa\u53ef\u80fd\u662f\uff1a\u201c\u89c6\u9891\u5c55\u793a\u4e86\u4e00\u573a\u70f9\u996a\u6bd4\u8d5b\uff0c\u9009\u624b\u5236\u4f5c\u4e86\u62ab\u8428\u3002\u201d<\/p>\n<h4>\u6587\u6863\u89e3\u6790<\/h4>\n<p>\u4e0a\u4f20 PDF \u6216\u56fe\u7247\uff0c\u7528\u7c7b\u4f3c\u56fe\u50cf\u8bc6\u522b\u7684\u4ee3\u7801\uff0c\u6539\u63d0\u793a\u4e3a\u201c\u63d0\u53d6\u8868\u683c\u6570\u636e\u201d\uff0c\u6a21\u578b\u4f1a\u8fd4\u56de\u7ed3\u6784\u5316\u7ed3\u679c\uff0c\u5982\uff1a<\/p>\n<pre><code>[{\"\u52171\": \"\u503c1\", \"\u52172\": \"\u503c2\"}]\r\n<\/code><\/pre>\n<h4>\u667a\u80fd\u4ee3\u7406<\/h4>\n<p>\u60f3\u63a7\u5236\u8bbe\u5907\uff1f\u7528\u00a0<code>cookbooks\/computer_use.ipynb<\/code>\u00a0\u793a\u4f8b\u3002\u8fd0\u884c\u540e\uff0c\u8f93\u5165\u201c\u6253\u5f00\u8bb0\u4e8b\u672c\u5e76\u8f93\u5165\u2018\u4f60\u597d\u2019\u201d\uff0c\u6a21\u578b\u4f1a\u6a21\u62df\u64cd\u4f5c\u5e76\u8fd4\u56de\u7ed3\u679c\u3002<\/p>\n<h3>Web \u754c\u9762\u4f7f\u7528<\/h3>\n<p>\u8fd0\u884c\u00a0<code>web_demo_mm.py<\/code>\u00a0\u542f\u52a8 Web \u754c\u9762\uff1a<\/p>\n<pre><code>python web_demo_mm.py\r\n<\/code><\/pre>\n<p>\u6d4f\u89c8\u5668\u8bbf\u95ee\u00a0<code>http:\/\/127.0.0.1:7860<\/code>\uff0c\u4e0a\u4f20\u6587\u4ef6\u6216\u8f93\u5165\u6307\u4ee4\uff0c\u6a21\u578b\u4f1a\u76f4\u63a5\u5904\u7406\u3002<\/p>\n<h4>\u4f18\u5316\u6027\u80fd<\/h4>\n<ul>\n<li><strong>Flash Attention 2<\/strong>\uff1a\u52a0\u901f\u63a8\u7406\uff0c\u5b89\u88c5\u540e\u7528\uff1a\n<pre><code>pip install -U flash-attn --no-build-isolation\r\npython web_demo_mm.py --flash-attn2\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u5206\u8fa8\u7387\u8c03\u6574<\/strong>\uff1a\u8bbe\u7f6e\u00a0<code>min_pixels<\/code>\u00a0\u548c\u00a0<code>max_pixels<\/code>\u00a0\u63a7\u5236\u56fe\u50cf\u5927\u5c0f\uff0c\u5982 256-1280 \u8303\u56f4\uff0c\u5e73\u8861\u901f\u5ea6\u548c\u5185\u5b58\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 \/>\n\u5b66\u751f\u4e0a\u4f20\u8bba\u6587\u56fe\u7247\uff0c\u6a21\u578b\u63d0\u53d6\u516c\u5f0f\u548c\u6570\u636e\uff0c\u751f\u6210\u5206\u6790\u62a5\u544a\u3002<\/li>\n<li><strong>\u89c6\u9891\u526a\u8f91<\/strong><br \/>\n\u521b\u4f5c\u8005\u8f93\u5165\u957f\u89c6\u9891\uff0c\u6a21\u578b\u63d0\u53d6\u5173\u952e\u7247\u6bb5\uff0c\u751f\u6210\u6458\u8981\u3002<\/li>\n<li><strong>\u4f01\u4e1a\u6587\u6863\u7ba1\u7406<\/strong><br \/>\n\u5458\u5de5\u4e0a\u4f20\u5408\u540c\u626b\u63cf\u4ef6\uff0c\u6a21\u578b\u63d0\u53d6\u6761\u6b3e\uff0c\u8f93\u51fa\u8868\u683c\u3002<\/li>\n<li><strong>\u667a\u80fd\u52a9\u624b<\/strong><br \/>\n\u7528\u6237\u7528\u56fe\u7247\u548c\u8bed\u97f3\u6307\u4ee4\uff0c\u8ba9\u6a21\u578b\u67e5\u8be2\u624b\u673a\u4e0a\u7684\u822a\u73ed\u4fe1\u606f\u3002<\/li>\n<\/ol>\n<p>&nbsp;<\/p>\n<h2>QA<\/h2>\n<ol>\n<li><strong>\u652f\u6301\u54ea\u4e9b\u8bed\u8a00\uff1f<\/strong><br \/>\n\u652f\u6301\u4e2d\u6587\u3001\u82f1\u6587\u53ca\u591a\u79cd\u8bed\u8a00\uff08\u5982\u6cd5\u8bed\u3001\u5fb7\u8bed\uff09\uff0c\u80fd\u8bc6\u522b\u591a\u8bed\u8a00\u6587\u5b57\u548c\u624b\u5199\u5185\u5bb9\u3002<\/li>\n<li><strong>\u786c\u4ef6\u8981\u6c42\u662f\u4ec0\u4e48\uff1f<\/strong><br \/>\n3B \u6a21\u578b\u9700 8GB \u663e\u5b58\uff0c7B \u9700 16GB\uff0c32B \u548c 72B \u5efa\u8bae 24GB+ \u663e\u5b58\u7684\u4e13\u4e1a\u8bbe\u5907\u3002<\/li>\n<li><strong>\u5982\u4f55\u5904\u7406\u8d85\u957f\u89c6\u9891\uff1f<\/strong><br \/>\n\u6a21\u578b\u7528\u52a8\u6001\u5e27\u7387\u91c7\u6837\uff0c\u80fd\u7406\u89e3\u6570\u5c0f\u65f6\u89c6\u9891\uff0c\u7cbe\u786e\u5b9a\u4f4d\u79d2\u7ea7\u4e8b\u4ef6\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>Qwen2.5-VL \u662f\u963f\u91cc\u5df4\u5df4\u4e91\uff08Alibaba Cloud\uff09Qwen \u56e2\u961f\u5f00\u53d1\u7684\u5f00\u6e90\u591a\u6a21\u6001\u5927\u6a21\u578b\u3002\u5b83\u80fd\u540c\u65f6\u5904\u7406\u6587\u672c\u3001\u56fe\u50cf\u3001\u89c6\u9891\u548c\u6587\u6863\uff0c\u662f Qwen2-VL \u7684\u5347\u7ea7\u7248\uff0c\u57fa\u4e8e Qwen2.5 \u8bed\u8a00\u6a21\u578b\u6784\u5efa\u3002\u5b98\u65b9\u79f0\uff0c\u5b83\u5728\u6587\u6863\u89e3\u6790\u3001\u89c6\u9891\u7406\u89e3\u548c\u667a&#8230;<\/p>\n","protected":false},"author":1,"featured_media":61744,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[230],"class_list":["post-29343","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","tag-aikaiyuanxiangmu"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/29343","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/comments?post=29343"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/29343\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/media\/61744"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/media?parent=29343"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/categories?post=29343"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/tags?post=29343"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}