\u5728\u73b0\u4ee3\u6587\u6863\u6570\u5b57\u5316\u5904\u7406\u4e2d，\u5149\u5b66\u5b57\u7b26\u8bc6\u522b（OCR）\u6280\u672f\u5df2\u7ecf\u6210\u4e3a\u4e0d\u53ef\u6216\u7f3a\u7684\u5de5\u5177\u3002\u4eca\u5929\u6211\u4eec\u8981\u4ecb\u7ecd\u7684DeepSeek-OCR-2，\u4e0d\u4ec5\u5177\u5907\u51fa\u8272\u7684\u6587\u5b57\u8bc6\u522b\u80fd\u529b，\u66f4\u7279\u522b\u7684\u662f\u5b83\u80fd\u591f\u4e0eNVIDIA Triton\u63a8\u7406\u670d\u52a1\u5668\u5b8c\u7f8e\u96c6\u6210，\u4e3a\u4f01\u4e1a\u7ea7\u5e94\u7528\u63d0\u4f9b\u9ad8\u6027\u80fd\u7684OCR\u670d\u52a1\u3002<\/p>\n
\u4f20\u7edf\u7684OCR\u90e8\u7f72\u5f80\u5f80\u9762\u4e34\u8fd9\u6837\u7684\u56f0\u5883：\u8981\u4e48\u90e8\u7f72\u7b80\u5355\u4f46\u6027\u80fd\u6709\u9650，\u8981\u4e48\u6027\u80fd\u5f3a\u5927\u4f46\u90e8\u7f72\u590d\u6742\u3002DeepSeek-OCR-2\u901a\u8fc7Triton\u63a8\u7406\u670d\u52a1\u5668\u7684\u96c6\u6210，\u5b8c\u7f8e\u89e3\u51b3\u4e86\u8fd9\u4e2a\u77db\u76fe\u3002\u65e0\u8bba\u4f60\u662f\u9700\u8981\u5904\u7406\u5927\u91cf\u6587\u6863\u7684\u4f01\u4e1a\u7528\u6237，\u8fd8\u662f\u9700\u8981\u9ad8\u8d28\u91cfOCR\u670d\u52a1\u7684\u5f00\u53d1\u8005，\u8fd9\u4e2a\u65b9\u6848\u90fd\u80fd\u63d0\u4f9b\u7a33\u5b9a\u53ef\u9760\u7684\u670d\u52a1\u3002<\/p>\n
\u5728\u672c\u6559\u7a0b\u4e2d，\u4f60\u5c06\u5b66\u4f1a\u5982\u4f55\u4ece\u96f6\u5f00\u59cb\u642d\u5efa\u8fd9\u5957OCR\u670d\u52a1\u7cfb\u7edf，\u5305\u62ec\u73af\u5883\u51c6\u5907\u3001\u6a21\u578b\u90e8\u7f72\u3001\u670d\u52a1\u914d\u7f6e\u548c\u5b9e\u9645\u6d4b\u8bd5\u3002\u6574\u4e2a\u8fc7\u7a0b\u6e05\u6670\u7b80\u5355，\u5373\u4f7f\u6ca1\u6709\u6df1\u539a\u7684\u673a\u5668\u5b66\u4e60\u80cc\u666f\u4e5f\u80fd\u8f7b\u677e\u4e0a\u624b\u3002<\/p>\n
\u5728\u5f00\u59cb\u90e8\u7f72\u4e4b\u524d，\u786e\u4fdd\u4f60\u7684\u7cfb\u7edf\u6ee1\u8db3\u4ee5\u4e0b\u57fa\u672c\u8981\u6c42：<\/p>\n
\u9996\u5148\u5b89\u88c5\u5fc5\u8981\u7684NVIDIA\u9a71\u52a8\u548cCUDA\u5de5\u5177\u5305：<\/p>\n
# \u6dfb\u52a0NVIDIA\u5305\u4ed3\u5e93
\nsudo apt update
\nsudo apt install apt-transport-https ca-certificates curl software-properties-common
\ncurl -fsSL https:\/\/nvidia.github.io\/nvidia-docker\/gpgkey | sudo apt-key add –
\ndistribution=$(. \/etc\/os-release;echo $ID$VERSION_ID)
\ncurl -s -L https:\/\/nvidia.github.io\/nvidia-docker\/$distribution\/nvidia-docker.list | sudo tee \/etc\/apt\/sources.list.d\/nvidia-docker.list<\/p>\n
# \u5b89\u88c5NVIDIA\u9a71\u52a8\u548cCUDA
\nsudo apt update
\nsudo apt install nvidia-driver-525 cuda-11-8<\/p>\n
\u5b89\u88c5\u5b8c\u6210\u540e，\u91cd\u542f\u7cfb\u7edf\u5e76\u9a8c\u8bc1\u5b89\u88c5：<\/p>\n
nvidia-smi # \u67e5\u770bGPU\u72b6\u6001
\nnvcc –version # \u9a8c\u8bc1CUDA\u5b89\u88c5<\/p>\n
Triton\u670d\u52a1\u5668\u901a\u5e38\u901a\u8fc7Docker\u5bb9\u5668\u8fd0\u884c：<\/p>\n
# \u5b89\u88c5Docker
\nsudo apt update
\nsudo apt install docker.io
\nsudo systemctl start docker
\nsudo systemctl enable docker<\/p>\n
# \u5b89\u88c5NVIDIA Container Toolkit
\ncurl -s -L https:\/\/nvidia.github.io\/nvidia-container-runtime\/gpgkey | sudo apt-key add –
\ndistribution=$(. \/etc\/os-release;echo $ID$VERSION_ID)
\ncurl -s -L https:\/\/nvidia.github.io\/nvidia-container-runtime\/$distribution\/nvidia-container-runtime.list | sudo tee \/etc\/apt\/sources.list.d\/nvidia-container-runtime.list
\nsudo apt update
\nsudo apt install nvidia-container-toolkit
\nsudo systemctl restart docker<\/p>\n
NVIDIA\u63d0\u4f9b\u4e86\u9884\u914d\u7f6e\u7684Triton\u670d\u52a1\u5668\u955c\u50cf，\u6211\u4eec\u76f4\u63a5\u4f7f\u7528\u5b98\u65b9\u955c\u50cf：<\/p>\n
# \u62c9\u53d6Triton\u670d\u52a1\u5668\u955c\u50cf
\ndocker pull nvcr.io\/nvidia\/tritonserver:23.07-py3<\/p>\n
# \u521b\u5efa\u6a21\u578b\u5b58\u50a8\u76ee\u5f55
\nmkdir -p triton_models\/deepseek-ocr-2\/1
\nmkdir -p triton_models\/deepseek-ocr-2\/config<\/p>\n
\u4e0b\u8f7dDeepSeek-OCR-2\u6a21\u578b\u6587\u4ef6\u5e76\u653e\u7f6e\u5230\u6b63\u786e\u4f4d\u7f6e：<\/p>\n
# \u4e0b\u8f7d\u6a21\u578b\u6587\u4ef6（\u8bf7\u66ff\u6362\u4e3a\u5b9e\u9645\u6a21\u578b\u4e0b\u8f7d\u94fe\u63a5）
\nwget -O triton_models\/deepseek-ocr-2\/1\/model.plan https:\/\/example.com\/deepseek-ocr-2\/model.plan
\nwget -O triton_models\/deepseek-ocr-2\/config\/config.pbtxt https:\/\/example.com\/deepseek-ocr-2\/config.pbtxt<\/p>\n
\u6a21\u578b\u914d\u7f6e\u6587\u4ef6\u793a\u4f8b（config.pbtxt）：<\/p>\n
name: "deepseek-ocr-2"
\nplatform: "tensorrt_plan"
\nmax_batch_size: 8
\ninput [
\n {
\n name: "input_image"
\n data_type: TYPE_UINT8
\n dims: [ -1, -1, 3 ]
\n }
\n]
\noutput [
\n {
\n name: "output_text"
\n data_type: TYPE_STRING
\n dims: [ -1 ]
\n }
\n]<\/p>\n
\u4f7f\u7528Docker\u542f\u52a8Triton\u63a8\u7406\u670d\u52a1\u5668：<\/p>\n
docker run -d –name triton-server \\\\
\n –gpus=all \\\\
\n -p 8000:8000 -p 8001:8001 -p 8002:8002 \\\\
\n -v $(pwd)\/triton_models:\/models \\\\
\n nvcr.io\/nvidia\/tritonserver:23.07-py3 \\\\
\n tritonserver –model-repository=\/models<\/p>\n
\u9a8c\u8bc1\u670d\u52a1\u5668\u662f\u5426\u6b63\u5e38\u542f\u52a8：<\/p>\n
# \u68c0\u67e5\u670d\u52a1\u5668\u72b6\u6001
\ncurl -v localhost:8000\/v2\/health\/ready<\/p>\n
# \u67e5\u770b\u5df2\u52a0\u8f7d\u6a21\u578b
\ncurl localhost:8000\/v2\/models\/deepseek-ocr-2<\/p>\n
\u4e3a\u4e86\u65b9\u4fbf\u6d4b\u8bd5，\u6211\u4eec\u5b89\u88c5Triton\u7684Python\u5ba2\u6237\u7aef：<\/p>\n
pip install tritonclient[all]<\/p>\n
\u521b\u5efa\u4e00\u4e2a\u7b80\u5355\u7684Python\u811a\u672c\u6765\u6d4b\u8bd5OCR\u670d\u52a1：<\/p>\n
import tritonclient.http as httpclient
\nimport cv2
\nimport numpy as np<\/p>\n
class DeepSeekOCRClient:
\n def __init__(self, url="localhost:8000"):
\n self.client = httpclient.InferenceServerClient(url=url)
\n self.model_name = "deepseek-ocr-2"<\/p>\n
def preprocess_image(self, image_path):
\n """\u9884\u5904\u7406\u8f93\u5165\u56fe\u50cf"""
\n image = cv2.imread(image_path)
\n image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
\n return image<\/p>\n
def ocr_inference(self, image_path):
\n """\u6267\u884cOCR\u63a8\u7406"""
\n # \u9884\u5904\u7406\u56fe\u50cf
\n image = self.preprocess_image(image_path)<\/p>\n
# \u51c6\u5907\u8f93\u5165
\n inputs = [httpclient.InferInput("input_image", image.shape, "UINT8")]
\n inputs[0].set_data_from_numpy(image)<\/p>\n
# \u51c6\u5907\u8f93\u51fa
\n outputs = [httpclient.InferRequestedOutput("output_text")]<\/p>\n
# \u53d1\u9001\u8bf7\u6c42
\n response = self.client.infer(
\n model_name=self.model_name,
\n inputs=inputs,
\n outputs=outputs
\n )<\/p>\n
# \u89e3\u6790\u7ed3\u679c
\n result = response.as_numpy("output_text")
\n return result[0].decode('utf-8')<\/p>\n
# \u6d4b\u8bd5\u4ee3\u7801
\nif __name__ == "__main__":
\n client = DeepSeekOCRClient()
\n result = client.ocr_inference("test_document.jpg")
\n print("\u8bc6\u522b\u7ed3\u679c:", result)<\/p>\n
\u5bf9\u4e8e\u9700\u8981\u5904\u7406\u5927\u91cf\u6587\u6863\u7684\u573a\u666f，\u53ef\u4ee5\u7f16\u5199\u6279\u91cf\u5904\u7406\u811a\u672c：<\/p>\n
import os
\nfrom concurrent.futures import ThreadPoolExecutor<\/p>\n
def process_document_batch(image_folder, output_folder):
\n """\u6279\u91cf\u5904\u7406\u6587\u6863\u56fe\u50cf"""
\n client = DeepSeekOCRClient()<\/p>\n
if not os.path.exists(output_folder):
\n os.makedirs(output_folder)<\/p>\n
def process_single(image_file):
\n try:
\n image_path = os.path.join(image_folder, image_file)
\n result = client.ocr_inference(image_path)<\/p>\n
# \u4fdd\u5b58\u7ed3\u679c
\n output_file = os.path.splitext(image_file)[0] + ".txt"
\n output_path = os.path.join(output_folder, output_file)<\/p>\n
with open(output_path, 'w', encoding='utf-8') as f:
\n f.write(result)<\/p>\n
print(f"\u5904\u7406\u5b8c\u6210: {image_file}")<\/p>\n
except Exception as e:
\n print(f"\u5904\u7406\u5931\u8d25 {image_file}: {str(e)}")<\/p>\n
# \u83b7\u53d6\u6240\u6709\u56fe\u50cf\u6587\u4ef6
\n image_files = [f for f in os.listdir(image_folder)
\n if f.lower().endswith(('.png', '.jpg', '.jpeg'))]<\/p>\n
# \u4f7f\u7528\u7ebf\u7a0b\u6c60\u5e76\u884c\u5904\u7406
\n with ThreadPoolExecutor(max_workers=4) as executor:
\n executor.map(process_single, image_files)<\/p>\n
# \u4f7f\u7528\u793a\u4f8b
\nprocess_document_batch("input_documents", "output_texts")<\/p>\n
\u6839\u636e\u4f60\u7684\u786c\u4ef6\u914d\u7f6e\u8c03\u6574Triton\u670d\u52a1\u5668\u53c2\u6570：<\/p>\n
# \u4f7f\u7528\u4f18\u5316\u53c2\u6570\u91cd\u65b0\u542f\u52a8\u670d\u52a1\u5668
\ndocker run -d –name triton-server-optimized \\\\
\n –gpus=all \\\\
\n -p 8000:8000 -p 8001:8001 -p 8002:8002 \\\\
\n -v $(pwd)\/triton_models:\/models \\\\
\n nvcr.io\/nvidia\/tritonserver:23.07-py3 \\\\
\n tritonserver –model-repository=\/models \\\\
\n –http-port=8000 \\\\
\n –grpc-port=8001 \\\\
\n –metrics-port=8002 \\\\
\n –model-control-mode=explicit \\\\
\n –load-model=deepseek-ocr-2 \\\\
\n –backend-config=tensorrt,load-timing-cache=1<\/p>\n
\u4f7f\u7528Prometheus\u548cGrafana\u76d1\u63a7\u670d\u52a1\u5668\u6027\u80fd：<\/p>\n
# prometheus.yml \u914d\u7f6e\u793a\u4f8b
\nglobal:
\n scrape_interval: 15s<\/p>\n
scrape_configs:
\n – job_name: 'triton'
\n static_configs:
\n – targets: ['localhost:8002']<\/p>\n
\u5bf9\u4e8e\u9ad8\u5e76\u53d1\u573a\u666f，\u53ef\u4ee5\u914d\u7f6e\u591a\u4e2aTriton\u5b9e\u4f8b\u5e76\u4f7f\u7528\u8d1f\u8f7d\u5747\u8861：<\/p>\n
from tritonclient import httpclient
\nimport random<\/p>\n
class LoadBalancedOCRClient:
\n def __init__(self, servers=["localhost:8000", "localhost:8003"]):
\n self.servers = servers
\n self.clients = [httpclient.InferenceServerClient(url=url) for url in servers]<\/p>\n
def get_client(self):
\n """\u968f\u673a\u9009\u62e9\u4e00\u4e2a\u53ef\u7528\u7684\u5ba2\u6237\u7aef"""
\n return random.choice(self.clients)<\/p>\n
def ocr_inference(self, image_path):
\n client = self.get_client()
\n # … \u5176\u4f59\u4ee3\u7801\u4e0e\u4e4b\u524d\u76f8\u540c<\/p>\n
\u95ee\u98981：GPU\u5185\u5b58\u4e0d\u8db3<\/p>\n
\u89e3\u51b3\u65b9\u6848：\u51cf\u5c11batch_size\u6216\u4f7f\u7528\u66f4\u5c0f\u7684\u6a21\u578b\u7248\u672c<\/p>\n
\u95ee\u98982：\u6a21\u578b\u52a0\u8f7d\u5931\u8d25<\/p>\n
\u89e3\u51b3\u65b9\u6848：\u68c0\u67e5\u6a21\u578b\u683c\u5f0f\u662f\u5426\u6b63\u786e，\u786e\u4fdd\u4f7f\u7528Triton\u652f\u6301\u7684\u683c\u5f0f<\/p>\n
\u95ee\u98983：\u63a8\u7406\u901f\u5ea6\u6162<\/p>\n
\u89e3\u51b3\u65b9\u6848：\u542f\u7528\u6a21\u578b\u4f18\u5316，\u4f7f\u7528TensorRT\u52a0\u901f<\/p>\n
# \u67e5\u770b\u670d\u52a1\u5668\u65e5\u5fd7
\ndocker logs triton-server<\/p>\n
# \u68c0\u67e5GPU\u72b6\u6001
\nnvidia-smi<\/p>\n
# \u6d4b\u8bd5\u670d\u52a1\u5668\u8fde\u901a\u6027
\ncurl -v localhost:8000\/v2\/health\/ready<\/p>\n
# \u67e5\u770b\u6a21\u578b\u72b6\u6001
\ncurl localhost:8000\/v2\/models\/deepseek-ocr-2<\/p>\n
\u901a\u8fc7\u672c\u6559\u7a0b，\u6211\u4eec\u6210\u529f\u5b8c\u6210\u4e86DeepSeek-OCR-2\u5728NVIDIA Triton\u63a8\u7406\u670d\u52a1\u5668\u4e0a\u7684\u90e8\u7f72\u548c\u96c6\u6210\u3002\u8fd9\u4e2a\u65b9\u6848\u4e0d\u4ec5\u63d0\u4f9b\u4e86\u9ad8\u6027\u80fd\u7684OCR\u670d\u52a1，\u8fd8\u5177\u5907\u4e86\u4f01\u4e1a\u7ea7\u5e94\u7528\u6240\u9700\u7684\u7a33\u5b9a\u6027\u3001\u53ef\u6269\u5c55\u6027\u548c\u6613\u7ef4\u62a4\u6027\u3002<\/p>\n
\u5173\u952e\u6536\u83b7：<\/p>\n
\u4e0b\u4e00\u6b65\u5efa\u8bae：<\/p>\n
\u8fd9\u5957\u90e8\u7f72\u65b9\u6848\u4e3a\u5904\u7406\u5927\u91cf\u6587\u6863\u8bc6\u522b\u4efb\u52a1\u63d0\u4f9b\u4e86\u53ef\u9760\u7684\u6280\u672f\u57fa\u7840，\u65e0\u8bba\u662f\u5b66\u672f\u7814\u7a76\u8fd8\u662f\u5546\u4e1a\u5e94\u7528，\u90fd\u80fd\u53d1\u6325\u91cd\u8981\u4f5c\u7528\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":"
DeepSeek-OCR-2\u90e8\u7f72\u6559\u7a0b：NVIDIA Triton\u63a8\u7406\u670d\u52a1\u5668\u96c6\u6210OCR\u670d\u52a1
\n1. \u5f15\u8a00：\u5f53OCR\u9047\u89c1\u9ad8\u6027\u80fd\u63a8\u7406
\n\u5728\u73b0\u4ee3\u6587\u6863\u6570\u5b57\u5316\u5904\u7406\u4e2d，\u5149\u5b66\u5b57\u7b26\u8bc6\u522b（OCR）\u6280\u672f\u5df2\u7ecf\u6210\u4e3a\u4e0d\u53ef\u6216\u7f3a\u7684\u5de5\u5177\u3002\u4eca\u5929\u6211\u4eec\u8981\u4ecb\u7ecd\u7684DeepSeek-OCR-2，\u4e0d\u4ec5\u5177\u5907\u51fa\u8272\u7684\u6587\u5b57\u8bc6\u522b\u80fd\u529b，\u66f4\u7279\u522b\u7684\u662f\u5b83\u80fd\u591f\u4e0eNVIDIA Triton\u63a8\u7406\u670d\u52a1\u5668\u5b8c\u7f8e\u96c6\u6210，\u4e3a\u4f01\u4e1a\u7ea7\u5e94\u7528\u63d0\u4f9b\u9ad8\u6027\u80fd\u7684O<\/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":[7591,756,86],"topic":[],"class_list":["post-85198","post","type-post","status-publish","format-standard","hentry","category-server","tag-ai","tag-ocr","tag-86"],"yoast_head":"\n