Hunyuan-MT-7B\u662f\u817e\u8baf\u6df7\u5143\u56e2\u961f\u57282025\u5e749\u6708\u5f00\u6e90\u7684\u591a\u8bed\u8a00\u7ffb\u8bd1\u6a21\u578b，\u62e5\u670970\u4ebf\u53c2\u6570，\u4e13\u95e8\u9488\u5bf933\u79cd\u8bed\u8a00\u7684\u53cc\u5411\u4e92\u8bd1\u4efb\u52a1\u8fdb\u884c\u4e86\u4f18\u5316\u3002\u8fd9\u4e2a\u6a21\u578b\u7279\u522b\u503c\u5f97\u5173\u6ce8\u7684\u662f，\u5b83\u4e0d\u4ec5\u652f\u6301\u4e3b\u6d41\u56fd\u9645\u8bed\u8a00，\u8fd8\u5305\u542b\u4e86\u85cf\u8bed\u3001\u8499\u53e4\u8bed\u3001\u7ef4\u543e\u5c14\u8bed\u3001\u54c8\u8428\u514b\u8bed\u3001\u671d\u9c9c\u8bed\u7b495\u79cd\u4e2d\u56fd\u5c11\u6570\u6c11\u65cf\u8bed\u8a00\u3002<\/p>\n
\u5728\u5b9e\u9645\u6d4b\u8bd5\u4e2d，Hunyuan-MT-7B\u8868\u73b0\u5353\u8d8a：\u5728WMT2025\u768431\u4e2a\u7ffb\u8bd1\u8d5b\u9053\u4e2d\u83b7\u5f97\u4e8630\u9879\u7b2c\u4e00，\u5728Flores-200\u57fa\u51c6\u6d4b\u8bd5\u4e2d，\u82f1\u8bd1\u591a\u8bed\u8a00\u8fbe\u523091.1%\u7684\u51c6\u786e\u7387，\u4e2d\u8bd1\u591a\u8bed\u8a00\u8fbe\u523087.6%\u7684\u51c6\u786e\u7387\u3002\u6700\u4ee4\u4eba\u60ca\u559c\u7684\u662f，\u4f7f\u7528BF16\u7cbe\u5ea6\u63a8\u7406\u65f6\u4ec5\u970016GB\u663e\u5b58，\u8ba9\u6d88\u8d39\u7ea7\u663e\u5361\u4e5f\u80fd\u6d41\u7545\u8fd0\u884c\u3002<\/p>\n
\u6838\u5fc3\u4f18\u52bf：<\/p>\n
\u5728\u5f00\u59cb\u90e8\u7f72\u524d，\u8bf7\u786e\u4fdd\u4f60\u7684\u7cfb\u7edf\u6ee1\u8db3\u4ee5\u4e0b\u6700\u4f4e\u8981\u6c42：<\/p>\n
\u786c\u4ef6\u8981\u6c42：<\/p>\n
\u8f6f\u4ef6\u8981\u6c42：<\/p>\n
\u672c\u6b21\u90e8\u7f72\u91c7\u7528vLLM + Open-WebUI\u7684\u7ec4\u5408\u65b9\u6848：<\/p>\n
\u5bf9\u4e8e\u5927\u591a\u6570\u7528\u6237，\u6211\u4eec\u63a8\u8350\u4f7f\u7528\u9884\u6784\u5efa\u7684Docker\u955c\u50cf\u8fdb\u884c\u5feb\u901f\u90e8\u7f72：<\/p>\n
# \u62c9\u53d6\u9884\u6784\u5efa\u7684\u955c\u50cf
\ndocker pull csdn-mirror\/hunyuan-mt-7b-vllm-webui:latest<\/p>\n
# \u8fd0\u884c\u5bb9\u5668
\ndocker run -d –gpus all \\\\
\n -p 7860:7860 \\\\
\n -p 8000:8000 \\\\
\n –name hunyuan-mt-7b \\\\
\n csdn-mirror\/hunyuan-mt-7b-vllm-webui:latest<\/p>\n
\u8fd9\u4e2a\u547d\u4ee4\u4f1a\u542f\u52a8\u4e00\u4e2a\u5305\u542b\u5b8c\u6574\u73af\u5883\u7684\u5bb9\u5668：<\/p>\n
\u5982\u679c\u4f60\u9700\u8981\u81ea\u5b9a\u4e49\u914d\u7f6e，\u53ef\u4ee5\u6309\u7167\u4ee5\u4e0b\u6b65\u9aa4\u624b\u52a8\u90e8\u7f72：<\/p>\n
# 1. \u514b\u9686\u9879\u76ee\u4ed3\u5e93
\ngit clone https:\/\/github.com\/Tencent\/HunyuanMT-7B.git
\ncd HunyuanMT-7B<\/p>\n
# 2. \u5b89\u88c5Python\u4f9d\u8d56
\npip install -r requirements.txt<\/p>\n
# 3. \u5b89\u88c5vLLM
\npip install vllm<\/p>\n
# 4. \u542f\u52a8vLLM\u670d\u52a1
\npython -m vllm.entrypoints.openai.api_server \\\\
\n –model Tencent\/HunyuanMT-7B-FP8 \\\\
\n –served-model-name hunyuan-mt-7b \\\\
\n –host 0.0.0.0 \\\\
\n –port 8000<\/p>\n
# 5. \u5b89\u88c5\u5e76\u542f\u52a8Open-WebUI
\ndocker run -d \\\\
\n -p 7860:7860 \\\\
\n -e OLLAMA_BASE_URL=http:\/\/localhost:8000 \\\\
\n –name open-webui \\\\
\n ghcr.io\/open-webui\/open-webui:main<\/p>\n
\u90e8\u7f72\u5b8c\u6210\u540e，\u53ef\u4ee5\u901a\u8fc7\u4ee5\u4e0b\u65b9\u5f0f\u9a8c\u8bc1\u670d\u52a1\u662f\u5426\u6b63\u5e38：<\/p>\n
# \u68c0\u67e5vLLM\u670d\u52a1
\ncurl http:\/\/localhost:8000\/v1\/models<\/p>\n
# \u9884\u671f\u8f93\u51fa\u7c7b\u4f3c：
\n# {"object":"list","data":[{"id":"hunyuan-mt-7b","object":"model",…}]}<\/p>\n
\u90e8\u7f72\u5b8c\u6210\u540e，\u7b49\u5f85\u51e0\u5206\u949f\u8ba9\u670d\u52a1\u5b8c\u5168\u542f\u52a8，\u7136\u540e\u901a\u8fc7\u6d4f\u89c8\u5668\u8bbf\u95ee：<\/p>\n
http:\/\/\u4f60\u7684\u670d\u52a1\u5668IP:7860<\/p>\n
\u4f7f\u7528\u4ee5\u4e0b\u6f14\u793a\u8d26\u53f7\u767b\u5f55：<\/p>\n
\u5728Open-WebUI\u754c\u9762\u4e2d，\u4f60\u53ef\u4ee5：<\/p>\n
\u9664\u4e86Web\u754c\u9762，\u4f60\u4e5f\u53ef\u4ee5\u901a\u8fc7API\u76f4\u63a5\u8c03\u7528\u7ffb\u8bd1\u670d\u52a1：<\/p>\n
import openai<\/p>\n
# \u914d\u7f6eAPI\u8fde\u63a5
\nclient = openai.OpenAI(
\n api_key="EMPTY",
\n base_url="http:\/\/localhost:8000\/v1"
\n)<\/p>\n
# \u6267\u884c\u7ffb\u8bd1\u4efb\u52a1
\nresponse = client.chat.completions.create(
\n model="hunyuan-mt-7b",
\n messages=[
\n {"role": "user", "content": "Translate this English text to Chinese: Hello, how are you today?"}
\n ],
\n temperature=0.7,
\n max_tokens=1000
\n)<\/p>\n
print(response.choices[0].message.content)<\/p>\n
NVIDIA Triton\u63a8\u7406\u670d\u52a1\u5668\u4e3a\u4f01\u4e1a\u7ea7\u90e8\u7f72\u63d0\u4f9b\u4e86\u591a\u9879\u4f18\u52bf：<\/p>\n
\u5728\u8fc1\u79fb\u5230Triton\u4e4b\u524d，\u9700\u8981\u5b8c\u6210\u4ee5\u4e0b\u51c6\u5907：<\/p>\n
# \u5b89\u88c5Triton\u5ba2\u6237\u7aef\u5e93
\npip install tritonclient[all]<\/p>\n
# \u5bfc\u51fa\u5f53\u524dvLLM\u6a21\u578b\u4e3aONNX\u683c\u5f0f（\u5982\u679c\u9700\u8981）
\npython -m vllm.entrypoints.openai.api_server \\\\
\n –model Tencent\/HunyuanMT-7B-FP8 \\\\
\n –export-format onnx \\\\
\n –output-dir .\/hunyuan-mt-7b-onnx<\/p>\n
\u521b\u5efaTriton\u6a21\u578b\u914d\u7f6e\u6587\u4ef6（model-config.pbtxt）：<\/p>\n
name: "hunyuan-mt-7b"
\nplatform: "onnxruntime_onnx"
\nmax_batch_size: 8
\ninput [
\n {
\n name: "input_ids"
\n data_type: TYPE_INT64
\n dims: [ -1 ]
\n }
\n]
\noutput [
\n {
\n name: "output"
\n data_type: TYPE_FP16
\n dims: [ -1, 51200 ]
\n }
\n]
\ninstance_group [
\n {
\n count: 1
\n kind: KIND_GPU
\n }
\n]<\/p>\n
# \u542f\u52a8Triton\u63a8\u7406\u670d\u52a1\u5668
\ndocker run -d –gpus=all \\\\
\n -p 8000:8000 -p 8001:8001 -p 8002:8002 \\\\
\n -v \/path\/to\/model\/repository:\/models \\\\
\n nvcr.io\/nvidia\/tritonserver:24.04-py3 \\\\
\n tritonserver –model-repository=\/models<\/p>\n
\u9488\u5bf9\u4e0d\u540c\u786c\u4ef6\u914d\u7f6e，\u63a8\u8350\u4ee5\u4e0b\u4f18\u5316\u65b9\u6848：<\/p>\n
RTX 4080 (16GB\u663e\u5b58)：<\/p>\n
# \u4f7f\u7528FP8\u91cf\u5316\u7248\u672c
\npython -m vllm.entrypoints.openai.api_server \\\\
\n –model Tencent\/HunyuanMT-7B-FP8 \\\\
\n –gpu-memory-utilization 0.9<\/p>\n
A100 (40GB\/80GB\u663e\u5b58)：<\/p>\n
# \u4f7f\u7528BF16\u539f\u7248，\u652f\u6301\u66f4\u5927\u6279\u91cf\u5904\u7406
\npython -m vllm.entrypoints.openai.api_server \\\\
\n –model Tencent\/HunyuanMT-7B \\\\
\n –dtype bfloat16 \\\\
\n –max-num-seqs 16<\/p>\n
\u901a\u8fc7\u8c03\u6574\u4ee5\u4e0b\u53c2\u6570\u53ef\u4ee5\u663e\u8457\u63d0\u5347\u63a8\u7406\u901f\u5ea6：<\/p>\n
# \u4f18\u5316vLLM\u914d\u7f6e
\npython -m vllm.entrypoints.openai.api_server \\\\
\n –model Tencent\/HunyuanMT-7B-FP8 \\\\
\n –max-model-len 8192 \\\\ # \u6839\u636e\u5b9e\u9645\u9700\u6c42\u8c03\u6574
\n –swap-space 4 \\\\ # GPU\u663e\u5b58\u4ea4\u6362\u7a7a\u95f4
\n –block-size 16 \\\\ # \u6ce8\u610f\u529b\u5757\u5927\u5c0f
\n –enable-prefix-caching # \u542f\u7528\u524d\u7f00\u7f13\u5b58\u52a0\u901f<\/p>\n
Q: \u90e8\u7f72\u540e\u65e0\u6cd5\u8bbf\u95eeWeb\u754c\u9762\u600e\u4e48\u529e？ A: \u9996\u5148\u68c0\u67e5\u670d\u52a1\u662f\u5426\u6b63\u5e38\u542f\u52a8：<\/p>\n
# \u68c0\u67e5\u5bb9\u5668\u72b6\u6001
\ndocker ps
\ndocker logs hunyuan-mt-7b<\/p>\n
# \u68c0\u67e5\u7aef\u53e3\u5360\u7528
\nnetstat -tlnp | grep 7860<\/p>\n
Q: \u663e\u5b58\u4e0d\u8db3\u5982\u4f55\u89e3\u51b3？ A: \u5c1d\u8bd5\u4ee5\u4e0b\u65b9\u6848：<\/p>\n
Q: \u652f\u6301\u54ea\u4e9b\u6587\u4ef6\u683c\u5f0f\u7684\u7ffb\u8bd1？ A: \u76ee\u524d\u4e3b\u8981\u901a\u8fc7\u6587\u672c\u8f93\u5165\u8fdb\u884c\u7ffb\u8bd1，\u652f\u6301：<\/p>\n
Q: \u957f\u6587\u6863\u7ffb\u8bd1\u5982\u4f55\u4fdd\u8bc1\u8d28\u91cf？ A: Hunyuan-MT-7B\u539f\u751f\u652f\u630132k token\u4e0a\u4e0b\u6587，\u4f46\u5bf9\u4e8e\u8d85\u957f\u6587\u6863，\u5efa\u8bae：<\/p>\n
Hunyuan-MT-7B\u4f5c\u4e3a\u4e00\u4e2a\u5f00\u6e90\u7684\u591a\u8bed\u8a00\u7ffb\u8bd1\u6a21\u578b，\u5728\u6027\u80fd\u3001\u8d44\u6e90\u9700\u6c42\u548c\u5546\u7528\u53cb\u597d\u5ea6\u65b9\u9762\u90fd\u8868\u73b0\u51fa\u8272\u3002\u901a\u8fc7vLLM + Open-WebUI\u7684\u90e8\u7f72\u65b9\u6848，\u6211\u4eec\u53ef\u4ee5\u5feb\u901f\u642d\u5efa\u4e00\u4e2a\u529f\u80fd\u5b8c\u6574\u7684\u7ffb\u8bd1\u670d\u52a1，\u540c\u65f6\u4e3a\u540e\u7eed\u8fc1\u79fb\u5230NVIDIA Triton\u63a8\u7406\u670d\u52a1\u5668\u505a\u597d\u4e86\u51c6\u5907\u3002<\/p>\n
\u5173\u952e\u4f18\u52bf\u56de\u987e：<\/p>\n
\u65e0\u8bba\u662f\u5b66\u672f\u7814\u7a76\u3001\u4f01\u4e1a\u5e94\u7528\u8fd8\u662f\u4e2a\u4eba\u9879\u76ee，Hunyuan-MT-7B\u90fd\u662f\u4e00\u4e2a\u503c\u5f97\u5c1d\u8bd5\u7684\u9ad8\u8d28\u91cf\u7ffb\u8bd1\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n
\u83b7\u53d6\u66f4\u591aAI\u955c\u50cf<\/p>\n
\u60f3\u63a2\u7d22\u66f4\u591aAI\u955c\u50cf\u548c\u5e94\u7528\u573a\u666f？\u8bbf\u95ee CSDN\u661f\u56fe\u955c\u50cf\u5e7f\u573a，\u63d0\u4f9b\u4e30\u5bcc\u7684\u9884\u7f6e\u955c\u50cf，\u8986\u76d6\u5927\u6a21\u578b\u63a8\u7406\u3001\u56fe\u50cf\u751f\u6210\u3001\u89c6\u9891\u751f\u6210\u3001\u6a21\u578b\u5fae\u8c03\u7b49\u591a\u4e2a\u9886\u57df，\u652f\u6301\u4e00\u952e\u90e8\u7f72\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
Hunyuan-MT-7B\u955c\u50cf\u90e8\u7f72：\u652f\u6301NVIDIA Triton\u63a8\u7406\u670d\u52a1\u5668\u65e0\u7f1d\u8fc1\u79fb\u65b9\u6848
\n1. \u9879\u76ee\u6982\u8ff0
\nHunyuan-MT-7B\u662f\u817e\u8baf\u6df7\u5143\u56e2\u961f\u57282025\u5e749\u6708\u5f00\u6e90\u7684\u591a\u8bed\u8a00\u7ffb\u8bd1\u6a21\u578b，\u62e5\u670970\u4ebf\u53c2\u6570，\u4e13\u95e8\u9488\u5bf933\u79cd\u8bed\u8a00\u7684\u53cc\u5411\u4e92\u8bd1\u4efb\u52a1\u8fdb\u884c\u4e86\u4f18\u5316\u3002\u8fd9\u4e2a\u6a21\u578b\u7279\u522b\u503c\u5f97\u5173\u6ce8\u7684\u662f，\u5b83\u4e0d\u4ec5\u652f\u6301\u4e3b\u6d41\u56fd\u9645\u8bed\u8a00，\u8fd8\u5305\u542b\u4e86\u85cf\u8bed\u3001\u8499\u53e4\u8bed\u3001\u7ef4\u543e\u5c14\u8bed\u3001\u54c8\u8428\u514b\u8bed\u3001\u671d\u9c9c\u8bed\u7b495\u79cd\u4e2d\u56fd\u5c11\u6570\u6c11\u65cf\u8bed\u8a00\u3002
\n\u5728\u5b9e\u9645\u6d4b\u8bd5\u4e2d，Huny<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[9185,9184,6851],"topic":[],"class_list":["post-82294","post","type-post","status-publish","format-standard","hentry","category-server","tag-hunyuan-mt-7b","tag-9184","tag-gpu"],"yoast_head":"\n