\u5982\u679c\u4f60\u7528\u8fc7\u6587\u751f\u56fe\u6a21\u578b，\u53ef\u80fd\u6709\u8fc7\u8fd9\u6837\u7684\u4f53\u9a8c：\u8f93\u5165\u4e00\u6bb5\u63cf\u8ff0，\u6ee1\u6000\u671f\u5f85\u5730\u7b49\u5f85，\u7136\u540e\u5f97\u5230\u4e00\u5f20\u7cbe\u7f8e\u7684\u56fe\u7247\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u5f88\u9177，\u4f46\u5f53\u4f60\u9700\u8981\u6279\u91cf\u751f\u6210\u56fe\u7247\u65f6，\u95ee\u9898\u5c31\u6765\u4e86\u2014\u2014\u4e00\u5f20\u4e00\u5f20\u5730\u751f\u6210，\u6548\u7387\u5b9e\u5728\u592a\u4f4e\u3002<\/p>\n
\u6700\u8fd1\u6211\u5728\u90e8\u7f72\u201c\u4e9a\u6d32\u7f8e\u5973-\u9020\u76f8Z-Turbo\u201d\u8fd9\u4e2a\u6587\u751f\u56fe\u6a21\u578b\u65f6，\u5c31\u9047\u5230\u4e86\u8fd9\u4e2a\u6311\u6218\u3002\u8fd9\u4e2a\u57fa\u4e8eZ-Image-Turbo LoRA\u7684\u6a21\u578b，\u5728\u751f\u6210\u4e9a\u6d32\u98ce\u683c\u4eba\u50cf\u65b9\u9762\u6548\u679c\u5f88\u4e0d\u9519，\u4f46\u9ed8\u8ba4\u914d\u7f6e\u4e0b，\u4e00\u6b21\u53ea\u80fd\u5904\u7406\u4e00\u4e2a\u8bf7\u6c42\u3002\u5bf9\u4e8e\u9700\u8981\u6279\u91cf\u751f\u6210\u56fe\u7247\u7684\u573a\u666f（\u6bd4\u5982\u7535\u5546\u5546\u54c1\u56fe\u3001\u793e\u4ea4\u5a92\u4f53\u914d\u56fe\u3001\u5185\u5bb9\u521b\u4f5c\u7b49），\u8fd9\u663e\u7136\u4e0d\u591f\u7528\u3002<\/p>\n
\u7ecf\u8fc7\u4e00\u756a\u6298\u817e，\u6211\u6210\u529f\u5728\u5355\u5f20A10G\u663e\u5361\u7684\u670d\u52a1\u5668\u4e0a，\u5c06\u8fd9\u4e2a\u6a21\u578b\u7684\u5e76\u53d1\u5904\u7406\u80fd\u529b\u63d0\u5347\u5230\u4e8610\u4e2a\u8bf7\u6c42\u4ee5\u4e0a\u3002\u8fd9\u610f\u5473\u7740\u4ec0\u4e48\u5462？\u7b80\u5355\u6765\u8bf4，\u539f\u6765\u751f\u621010\u5f20\u56fe\u7247\u53ef\u80fd\u9700\u898110\u5206\u949f，\u73b0\u5728\u53ef\u80fd\u53ea\u9700\u89811-2\u5206\u949f\u3002\u6548\u7387\u7684\u63d0\u5347\u662f\u5b9e\u5b9e\u5728\u5728\u7684\u3002<\/p>\n
\u8fd9\u7bc7\u6587\u7ae0，\u6211\u5c31\u6765\u5206\u4eab\u4e00\u4e0b\u5177\u4f53\u7684\u8c03\u4f18\u65b9\u6cd5\u548c\u5b9e\u8df5\u7ecf\u9a8c\u3002\u65e0\u8bba\u4f60\u662f\u521a\u63a5\u89e6\u6a21\u578b\u90e8\u7f72\u7684\u65b0\u624b，\u8fd8\u662f\u6709\u4e00\u5b9a\u7ecf\u9a8c\u4f46\u60f3\u63d0\u5347\u670d\u52a1\u6027\u80fd\u7684\u5f00\u53d1\u8005，\u76f8\u4fe1\u90fd\u80fd\u4ece\u4e2d\u83b7\u5f97\u4e00\u4e9b\u5b9e\u7528\u7684\u601d\u8def\u3002<\/p>\n
\u5728\u5f00\u59cb\u6027\u80fd\u8c03\u4f18\u4e4b\u524d，\u6211\u4eec\u5148\u786e\u4fdd\u57fa\u7840\u73af\u5883\u5df2\u7ecf\u6b63\u786e\u90e8\u7f72\u3002\u8fd9\u91cc\u4f7f\u7528\u7684\u662fXinference\u6846\u67b6\u6765\u90e8\u7f72\u6a21\u578b，\u5e76\u7528Gradio\u642d\u5efa\u4e86\u4e00\u4e2a\u7b80\u5355\u7684Web\u754c\u9762\u3002<\/p>\n
\u9996\u5148，\u786e\u8ba4\u4f60\u7684\u670d\u52a1\u5668\u914d\u7f6e\u3002\u6211\u4f7f\u7528\u7684\u662f\u5355\u5361A10G\u670d\u52a1\u5668，\u5177\u4f53\u914d\u7f6e\u5982\u4e0b：<\/p>\n
A10G\u663e\u5361\u768424GB\u663e\u5b58\u5bf9\u4e8e\u6587\u751f\u56fe\u6a21\u578b\u6765\u8bf4\u662f\u6bd4\u8f83\u5145\u88d5\u7684，\u8fd9\u4e3a\u6211\u4eec\u540e\u7eed\u7684\u5e76\u53d1\u4f18\u5316\u63d0\u4f9b\u4e86\u786c\u4ef6\u57fa\u7840\u3002<\/p>\n
\u6309\u7167\u6807\u51c6\u7684\u90e8\u7f72\u6d41\u7a0b，\u542f\u52a8Xinference\u670d\u52a1\u540e，\u6211\u4eec\u9700\u8981\u786e\u8ba4\u6a21\u578b\u662f\u5426\u52a0\u8f7d\u6210\u529f\u3002\u53ef\u4ee5\u901a\u8fc7\u67e5\u770b\u65e5\u5fd7\u6587\u4ef6\u6765\u786e\u8ba4：<\/p>\n
# \u67e5\u770b\u670d\u52a1\u542f\u52a8\u65e5\u5fd7
\ncat \/root\/workspace\/xinference.log<\/p>\n
\u5982\u679c\u770b\u5230\u7c7b\u4f3c\u4e0b\u9762\u7684\u8f93\u51fa，\u8bf4\u660e\u6a21\u578b\u670d\u52a1\u5df2\u7ecf\u6210\u529f\u542f\u52a8：<\/p>\n
INFO: Model 'asian-beauty-z-turbo' loaded successfully
\nINFO: GPU memory allocated: 8.2GB
\nINFO: Service started on port 9997<\/p>\n
\u901a\u8fc7Gradio\u7684Web\u754c\u9762，\u6211\u4eec\u53ef\u4ee5\u5148\u6d4b\u8bd5\u4e00\u4e0b\u57fa\u7840\u7684\u5355\u8bf7\u6c42\u751f\u6210\u529f\u80fd\u3002\u8f93\u5165\u4e00\u6bb5\u63cf\u8ff0，\u6bd4\u5982\u201c\u4e00\u4f4d\u5e74\u8f7b\u7684\u4e9a\u6d32\u5973\u6027，\u957f\u53d1，\u5728\u6a31\u82b1\u6811\u4e0b\u5fae\u7b11，\u9633\u5149\u660e\u5a9a\u201d，\u70b9\u51fb\u751f\u6210\u6309\u94ae\u3002<\/p>\n
\u5982\u679c\u4e00\u5207\u6b63\u5e38，\u4f60\u5e94\u8be5\u80fd\u5728\u51e0\u79d2\u5230\u5341\u51e0\u79d2\u5185\u770b\u5230\u751f\u6210\u7684\u56fe\u7247\u3002\u8fd9\u4e2a\u54cd\u5e94\u65f6\u95f4\u662f\u6211\u4eec\u540e\u7eed\u4f18\u5316\u7684\u57fa\u51c6\u2014\u2014\u6211\u4eec\u8981\u5728\u4fdd\u6301\u56fe\u7247\u8d28\u91cf\u7684\u524d\u63d0\u4e0b，\u5c3d\u53ef\u80fd\u7f29\u77ed\u8fd9\u4e2a\u65f6\u95f4，\u5e76\u4e14\u652f\u6301\u540c\u65f6\u5904\u7406\u591a\u4e2a\u8bf7\u6c42\u3002<\/p>\n
\u5728\u5f00\u59cb\u5177\u4f53\u7684\u8c03\u4f18\u64cd\u4f5c\u4e4b\u524d，\u6211\u4eec\u9700\u8981\u5148\u7406\u89e3\u6a21\u578b\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u6027\u80fd\u74f6\u9888\u5728\u54ea\u91cc\u3002\u8fd9\u5bf9\u4e8e\u5236\u5b9a\u6709\u6548\u7684\u4f18\u5316\u7b56\u7565\u81f3\u5173\u91cd\u8981\u3002<\/p>\n
\u901a\u8fc7\u76d1\u63a7\u5de5\u5177（\u5982nvidia-smi\u3001htop\u7b49）\u89c2\u5bdf\u6a21\u578b\u8fd0\u884c\u65f6\u7684\u8d44\u6e90\u4f7f\u7528\u60c5\u51b5，\u6211\u53d1\u73b0\u4e86\u51e0\u4e2a\u5173\u952e\u95ee\u9898：<\/p>\n
\u6587\u751f\u56fe\u6a21\u578b\u7684\u5e76\u53d1\u5904\u7406\u5e76\u4e0d\u662f\u7b80\u5355\u7684\u540c\u65f6\u8fd0\u884c\u591a\u4e2a\u6a21\u578b\u5b9e\u4f8b\u3002\u4e3b\u8981\u6311\u6218\u5305\u62ec：<\/p>\n
\u57fa\u4e8e\u4ee5\u4e0a\u5206\u6790，\u6211\u5236\u5b9a\u4e86\u660e\u786e\u7684\u4f18\u5316\u76ee\u6807：<\/p>\n
\u63a5\u4e0b\u6765，\u6211\u4eec\u8fdb\u5165\u5177\u4f53\u7684\u8c03\u4f18\u64cd\u4f5c\u3002\u8fd9\u4e9b\u65b9\u6cd5\u90fd\u662f\u7ecf\u8fc7\u5b9e\u9645\u6d4b\u8bd5\u9a8c\u8bc1\u6709\u6548\u7684，\u4f60\u53ef\u4ee5\u6839\u636e\u81ea\u5df1\u7684\u73af\u5883\u8fdb\u884c\u8c03\u6574\u3002<\/p>\n
\u9ed8\u8ba4\u7684\u6a21\u578b\u52a0\u8f7d\u65b9\u5f0f\u53ef\u80fd\u4e0d\u662f\u6700\u4f18\u7684\u3002\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u8c03\u6574\u52a0\u8f7d\u53c2\u6570\u6765\u63d0\u5347\u6027\u80fd：<\/p>\n
# \u4f18\u5316\u540e\u7684\u6a21\u578b\u52a0\u8f7d\u914d\u7f6e
\nmodel_config = {
\n "model_name": "asian-beauty-z-turbo",
\n "model_format": "safetensors",
\n "device": "cuda",
\n # \u542f\u7528\u66f4\u5feb\u7684\u6ce8\u610f\u529b\u673a\u5236
\n "enable_xformers": True,
\n # \u4f7f\u7528\u534a\u7cbe\u5ea6\u6d6e\u70b9\u6570，\u51cf\u5c11\u663e\u5b58\u5360\u7528
\n "dtype": "float16",
\n # \u9884\u5206\u914d\u663e\u5b58，\u51cf\u5c11\u8fd0\u884c\u65f6\u5206\u914d\u5f00\u9500
\n "max_memory": 0.9, # \u4f7f\u752890%\u7684\u663e\u5b58
\n}<\/p>\n
\u5173\u952e\u4f18\u5316\u70b9：<\/p>\n
\u8fd9\u662f\u63d0\u5347\u5e76\u53d1\u80fd\u529b\u7684\u5173\u952e\u6280\u672f\u3002\u6211\u4eec\u4e0d\u662f\u7b80\u5355\u5730\u4e3a\u6bcf\u4e2a\u8bf7\u6c42\u542f\u52a8\u4e00\u4e2a\u72ec\u7acb\u7684\u63a8\u7406\u8fc7\u7a0b，\u800c\u662f\u5c06\u591a\u4e2a\u8bf7\u6c42\u667a\u80fd\u5730\u7ec4\u5408\u6210\u4e00\u4e2a\u6279\u6b21\u8fdb\u884c\u5904\u7406\u3002<\/p>\n
class DynamicBatchProcessor:
\n def __init__(self, max_batch_size=8, timeout=0.1):
\n """
\n \u52a8\u6001\u6279\u6b21\u5904\u7406\u5668<\/p>\n
\u53c2\u6570：
\n – max_batch_size: \u6700\u5927\u6279\u6b21\u5927\u5c0f，\u6839\u636e\u663e\u5b58\u8c03\u6574
\n – timeout: \u7b49\u5f85\u65b0\u8bf7\u6c42\u7684\u6700\u5927\u65f6\u95f4（\u79d2）
\n """
\n self.max_batch_size = max_batch_size
\n self.timeout = timeout
\n self.request_queue = []
\n self.processing_lock = threading.Lock()<\/p>\n
def add_request(self, prompt, callback):
\n """\u6dfb\u52a0\u751f\u6210\u8bf7\u6c42\u5230\u961f\u5217"""
\n with self.processing_lock:
\n self.request_queue.append({
\n 'prompt': prompt,
\n 'callback': callback,
\n 'timestamp': time.time()
\n })<\/p>\n
def process_batch(self):
\n """\u5904\u7406\u4e00\u4e2a\u6279\u6b21\u7684\u8bf7\u6c42"""
\n with self.processing_lock:
\n if not self.request_queue:
\n return<\/p>\n
# \u83b7\u53d6\u5f53\u524d\u961f\u5217\u4e2d\u7684\u6240\u6709\u8bf7\u6c42
\n current_batch = self.request_queue[:self.max_batch_size]
\n self.request_queue = self.request_queue[self.max_batch_size:]<\/p>\n
if not current_batch:
\n return<\/p>\n
# \u63d0\u53d6\u6240\u6709prompt
\n prompts = [req['prompt'] for req in current_batch]<\/p>\n
# \u6279\u91cf\u751f\u6210\u56fe\u7247
\n try:
\n images = self.batch_generate(prompts)<\/p>\n
# \u56de\u8c03\u5904\u7406\u7ed3\u679c
\n for i, req in enumerate(current_batch):
\n if i < len(images):
\n req['callback'](images[i])
\n else:
\n req['callback'](None)
\n except Exception as e:
\n # \u9519\u8bef\u5904\u7406
\n for req in current_batch:
\n req['callback'](None, str(e))<\/p>\n
def batch_generate(self, prompts):
\n """\u5b9e\u9645\u7684\u6279\u91cf\u751f\u6210\u51fd\u6570"""
\n # \u8fd9\u91cc\u8c03\u7528\u6a21\u578b\u7684\u6279\u91cf\u751f\u6210\u63a5\u53e3
\n # \u5177\u4f53\u5b9e\u73b0\u53d6\u51b3\u4e8e\u4f60\u4f7f\u7528\u7684\u63a8\u7406\u6846\u67b6
\n pass<\/p>\n
\u8fd9\u4e2a\u52a8\u6001\u6279\u6b21\u5904\u7406\u5668\u7684\u6838\u5fc3\u601d\u60f3\u662f：<\/p>\n
\u8981\u652f\u6301\u66f4\u591a\u5e76\u53d1，\u5fc5\u987b\u7cbe\u7ec6\u5316\u7ba1\u7406\u663e\u5b58\u4f7f\u7528\u3002\u4ee5\u4e0b\u662f\u4e00\u4e9b\u6709\u6548\u7684\u7b56\u7565：<\/p>\n
# \u663e\u5b58\u4f18\u5316\u914d\u7f6e
\nmemory_config = {
\n # \u542f\u7528\u663e\u5b58\u6c60，\u51cf\u5c11\u788e\u7247
\n "enable_memory_pool": True,
\n # \u8bbe\u7f6e\u7f13\u5b58\u5927\u5c0f，\u52a0\u901f\u91cd\u590d\u8bf7\u6c42
\n "cache_size": 5, # \u7f13\u5b58\u6700\u8fd15\u4e2a\u63d0\u793a\u8bcd\u7684\u4e2d\u95f4\u7ed3\u679c
\n # \u4f7f\u7528\u68af\u5ea6\u68c0\u67e5\u70b9，\u7528\u8ba1\u7b97\u65f6\u95f4\u6362\u663e\u5b58
\n "gradient_checkpointing": True,
\n # \u6e05\u7406\u7b56\u7565：\u5b9a\u671f\u6e05\u7406\u4e0d\u7528\u7684\u7f13\u5b58
\n "cleanup_interval": 60, # \u6bcf60\u79d2\u6e05\u7406\u4e00\u6b21
\n}<\/p>\n
\u5177\u4f53\u4f18\u5316\u63aa\u65bd：<\/p>\n
\u5f53\u5e76\u53d1\u8bf7\u6c42\u5f88\u591a\u65f6，\u9700\u8981\u4e00\u4e2a\u667a\u80fd\u7684\u8bf7\u6c42\u8c03\u5ea6\u7cfb\u7edf：<\/p>\n
class RequestScheduler:
\n def __init__(self, max_concurrent=10):
\n self.max_concurrent = max_concurrent
\n self.active_requests = 0
\n self.waiting_queue = []
\n self.completed_requests = 0<\/p>\n
def submit_request(self, prompt, priority=1):
\n """\u63d0\u4ea4\u751f\u6210\u8bf7\u6c42"""
\n # \u6839\u636e\u5f53\u524d\u8d1f\u8f7d\u51b3\u5b9a\u662f\u5426\u7acb\u5373\u5904\u7406
\n if self.active_requests < self.max_concurrent:
\n self.active_requests += 1
\n return self.process_immediately(prompt)
\n else:
\n # \u52a0\u5165\u7b49\u5f85\u961f\u5217，\u6309\u4f18\u5148\u7ea7\u6392\u5e8f
\n self.waiting_queue.append({
\n 'prompt': prompt,
\n 'priority': priority,
\n 'submit_time': time.time()
\n })
\n self.waiting_queue.sort(key=lambda x: x['priority'], reverse=True)
\n return self.create_pending_response()<\/p>\n
def request_completed(self):
\n """\u8bf7\u6c42\u5b8c\u6210\u56de\u8c03"""
\n self.active_requests -= 1
\n self.completed_requests += 1<\/p>\n
# \u68c0\u67e5\u662f\u5426\u6709\u7b49\u5f85\u7684\u8bf7\u6c42
\n if self.waiting_queue and self.active_requests < self.max_concurrent:
\n next_request = self.waiting_queue.pop(0)
\n self.process_waiting_request(next_request)<\/p>\n
\u8fd9\u4e2a\u8c03\u5ea6\u5668\u7684\u4e3b\u8981\u529f\u80fd：<\/p>\n
\u7406\u8bba\u8bf4\u5b8c\u4e86，\u6211\u4eec\u6765\u770b\u770b\u5b9e\u9645\u6548\u679c\u3002\u4ee5\u4e0b\u662f\u6211\u5728A10G\u670d\u52a1\u5668\u4e0a\u7684\u6d4b\u8bd5\u7ed3\u679c\u548c\u914d\u7f6e\u3002<\/p>\n
\u7ecf\u8fc7\u591a\u6b21\u6d4b\u8bd5\u548c\u8c03\u6574，\u6211\u6700\u7ec8\u91c7\u7528\u7684\u914d\u7f6e\u5982\u4e0b：<\/p>\n
# deployment-config.yaml
\nmodel_config:
\n name: "asian-beauty-z-turbo"
\n dtype: "float16"
\n enable_xformers: true
\n enable_attention_slicing: true<\/p>\n
inference_config:
\n batch_size: 4 # \u6bcf\u4e2a\u6279\u6b21\u7684\u56fe\u7247\u6570\u91cf
\n num_inference_steps: 30 # \u63a8\u7406\u6b65\u6570，\u5e73\u8861\u901f\u5ea6\u548c\u8d28\u91cf
\n guidance_scale: 7.5 # \u6307\u5bfc\u5c3a\u5ea6<\/p>\n
performance_config:
\n max_concurrent_requests: 12 # \u6700\u5927\u5e76\u53d1\u8bf7\u6c42\u6570
\n max_batch_size: 4 # \u6700\u5927\u6279\u6b21\u5927\u5c0f
\n request_timeout: 30 # \u8bf7\u6c42\u8d85\u65f6\u65f6\u95f4（\u79d2）
\n enable_memory_pool: true
\n cache_enabled: true
\n cache_size: 10<\/p>\n
monitoring_config:
\n enable_metrics: true
\n metrics_port: 9090
\n log_level: "INFO"<\/p>\n
\u6211\u8bbe\u8ba1\u4e86\u51e0\u7ec4\u6d4b\u8bd5\u6765\u9a8c\u8bc1\u4f18\u5316\u6548\u679c：<\/p>\n
\u6d4b\u8bd51：\u5355\u8bf7\u6c42\u54cd\u5e94\u65f6\u95f4\u5bf9\u6bd4<\/p>\n
| \u4f18\u5316\u524d（\u5355\u8bf7\u6c42）<\/td>\n | 8.2\u79d2<\/td>\n | 9.5GB<\/td>\n | 45%<\/td>\n<\/tr>\n | ||||||||||||||||||||||||||||
| \u4f18\u5316\u540e（\u5355\u8bf7\u6c42）<\/td>\n | 6.5\u79d2<\/td>\n | 8.1GB<\/td>\n | 52%<\/td>\n<\/tr>\n | ||||||||||||||||||||||||||||
| \u63d0\u5347\u6548\u679c<\/td>\n | \u51cf\u5c1120.7%<\/td>\n | \u51cf\u5c1114.7%<\/td>\n | \u63d0\u534715.6%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u6d4b\u8bd52：\u5e76\u53d1\u5904\u7406\u80fd\u529b\u6d4b\u8bd5<\/p>\n
|