\u6700\u8fd1，Google\u5f00\u6e90\u7684Gemma-3-12B-IT\u6a21\u578b\u5728\u5f00\u53d1\u8005\u793e\u533a\u5f15\u8d77\u4e86\u4e0d\u5c0f\u7684\u5173\u6ce8\u3002\u4f5c\u4e3a\u4e00\u4e2a120\u4ebf\u53c2\u6570\u7684\u6307\u4ee4\u5fae\u8c03\u6a21\u578b，\u5b83\u5728\u63a8\u7406\u80fd\u529b\u3001\u591a\u8bed\u8a00\u652f\u6301\u548c\u6548\u7387\u4e0a\u76f8\u6bd4\u524d\u4ee3\u6709\u4e86\u663e\u8457\u63d0\u5347\u3002\u4f46\u6700\u5438\u5f15\u6211\u7684，\u662f\u5b83\u5ba3\u79f0\u7684\u201c\u8f7b\u91cf\u7ea7\u201d\u7279\u6027\u2014\u2014\u4e00\u4e2a23GB\u7684\u6a21\u578b，\u771f\u7684\u80fd\u572832GB\u5185\u5b58\u7684\u670d\u52a1\u5668\u4e0a\u7a33\u5b9a\u8fd0\u884c\u5417？<\/p>\n
\u8fd9\u6b63\u662f\u6211\u51b3\u5b9a\u8fdb\u884c\u8fd9\u6b21\u6d4b\u8bd5\u7684\u539f\u56e0\u3002\u5f88\u591a\u4e2d\u5c0f\u56e2\u961f\u548c\u4e2a\u4eba\u5f00\u53d1\u8005，\u624b\u5934\u5e76\u6ca1\u6709\u52a8\u8f84\u4e0a\u767eGB\u5185\u5b58\u7684\u8c6a\u534e\u914d\u7f6e\u3002\u5982\u679cGemma-3-12B-IT\u771f\u7684\u80fd\u572832GB\u5185\u5b58\u73af\u5883\u4e0b\u7a33\u5b9a\u5de5\u4f5c，\u90a3\u65e0\u7591\u4e3a\u66f4\u591a\u5f00\u53d1\u8005\u6253\u5f00\u4e86\u672c\u5730\u90e8\u7f72\u5927\u6a21\u578b\u7684\u5927\u95e8\u3002<\/p>\n
\u5728\u63a5\u4e0b\u6765\u7684\u5185\u5bb9\u91cc，\u6211\u5c06\u5206\u4eab\u4ece\u73af\u5883\u51c6\u5907\u5230\u538b\u529b\u6d4b\u8bd5\u7684\u5b8c\u6574\u8fc7\u7a0b，\u5305\u62ec\u9047\u5230\u7684\u5751\u3001\u89e3\u51b3\u7684\u65b9\u6848，\u4ee5\u53ca\u6700\u7ec8\u7684\u6027\u80fd\u8868\u73b0\u3002\u65e0\u8bba\u4f60\u662f\u60f3\u5728\u81ea\u5df1\u7684\u670d\u52a1\u5668\u4e0a\u90e8\u7f72，\u8fd8\u662f\u5355\u7eaf\u597d\u5947\u5927\u6a21\u578b\u5728\u6709\u9650\u8d44\u6e90\u4e0b\u7684\u8868\u73b0，\u8fd9\u7bc7\u6587\u7ae0\u90fd\u4f1a\u7ed9\u4f60\u4e00\u4e2a\u6e05\u6670\u7684\u7b54\u6848\u3002<\/p>\n
\u5148\u6765\u770b\u770b\u8fd9\u6b21\u6d4b\u8bd5\u7684\u786c\u4ef6\u6761\u4ef6，\u8fd9\u76f4\u63a5\u51b3\u5b9a\u4e86\u6a21\u578b\u7684\u8fd0\u884c\u4e0a\u9650：<\/p>\n
\u670d\u52a1\u5668\u914d\u7f6e\u6e05\u5355：<\/p>\n
\u8fd9\u4e2a\u914d\u7f6e\u5728\u4eca\u5929\u7684\u6807\u51c6\u4e0b\u7b97\u4e0d\u4e0a\u9ad8\u7aef，\u4f46\u6070\u6070\u4ee3\u8868\u4e86\u76f8\u5f53\u4e00\u90e8\u5206\u5f00\u53d1\u8005\u548c\u4e2d\u5c0f\u4f01\u4e1a\u7684\u5b9e\u9645\u60c5\u51b5\u3002\u6ca1\u6709\u72ec\u7acbGPU，\u5168\u9760CPU\u548c\u5185\u5b58\u786c\u625b\u3002<\/p>\n
\u5728\u5f00\u59cb\u90e8\u7f72\u4e4b\u524d，\u9700\u8981\u786e\u4fdd\u7cfb\u7edf\u73af\u5883\u51c6\u5907\u5c31\u7eea\u3002\u4ee5\u4e0b\u662f\u5173\u952e\u6b65\u9aa4：<\/p>\n
# \u66f4\u65b0\u7cfb\u7edf\u5305
\nsudo apt update && sudo apt upgrade -y<\/p>\n
# \u5b89\u88c5Python 3.11（Gemma-3\u63a8\u8350\u7248\u672c）
\nsudo apt install python3.11 python3.11-venv python3.11-dev -y<\/p>\n
# \u8bbe\u7f6ePython 3.11\u4e3a\u9ed8\u8ba4\u7248\u672c
\nsudo update-alternatives –install \/usr\/bin\/python3 python3 \/usr\/bin\/python3.11 1<\/p>\n
# \u5b89\u88c5\u5fc5\u8981\u7684\u7cfb\u7edf\u4f9d\u8d56
\nsudo apt install build-essential cmake git wget curl -y<\/p>\n
# \u521b\u5efa\u4e13\u7528\u76ee\u5f55
\nmkdir -p ~\/ai-projects\/gemma-3
\ncd ~\/ai-projects\/gemma-3<\/p>\n
Gemma-3-12B-IT\u6a21\u578b\u6587\u4ef6\u5927\u7ea623GB，\u4e0b\u8f7d\u9700\u8981\u4e00\u4e9b\u65f6\u95f4\u548c\u7a7a\u95f4\u51c6\u5907：<\/p>\n
# \u521b\u5efa\u6a21\u578b\u5b58\u50a8\u76ee\u5f55
\nmkdir -p ~\/ai-models\/LLM-Research\/gemma-3-12b-it<\/p>\n
# \u4f7f\u7528wget\u4e0b\u8f7d（\u9700\u8981\u63d0\u524d\u83b7\u53d6\u4e0b\u8f7d\u94fe\u63a5）
\n# \u8fd9\u91cc\u5047\u8bbe\u4f60\u5df2\u7ecf\u4eceHugging Face\u6216\u5b98\u65b9\u6e20\u9053\u83b7\u5f97\u4e86\u4e0b\u8f7d\u6743\u9650
\n# wget -c [\u6a21\u578b\u4e0b\u8f7d\u94fe\u63a5] -O ~\/ai-models\/LLM-Research\/gemma-3-12b-it\/model.bin<\/p>\n
# \u6216\u8005\u4f7f\u7528git lfs（\u5982\u679c\u6a21\u578b\u6258\u7ba1\u5728Git\u4e0a）
\n# git lfs install
\n# git clone [\u6a21\u578b\u4ed3\u5e93\u5730\u5740] ~\/ai-models\/LLM-Research\/gemma-3-12b-it<\/p>\n
\u91cd\u8981\u63d0\u793a：\u7531\u4e8e\u6a21\u578b\u6587\u4ef6\u8f83\u5927，\u5efa\u8bae\u5728\u7f51\u7edc\u72b6\u51b5\u826f\u597d\u7684\u65f6\u5019\u4e0b\u8f7d，\u6216\u8005\u4f7f\u7528\u6709\u65ad\u70b9\u7eed\u4f20\u529f\u80fd\u7684\u5de5\u5177\u3002\u4e0b\u8f7d\u5b8c\u6210\u540e，\u8bb0\u5f97\u9a8c\u8bc1\u6587\u4ef6\u5b8c\u6574\u6027\u3002<\/p>\n
\u6211\u9009\u62e9\u4e86\u4e00\u4e2a\u5f00\u6e90\u7684Web UI\u9879\u76ee\u6765\u90e8\u7f72Gemma-3-12B-IT，\u8fd9\u6837\u53ef\u4ee5\u901a\u8fc7\u6d4f\u89c8\u5668\u76f4\u63a5\u4e0e\u6a21\u578b\u4ea4\u4e92，\u6d4b\u8bd5\u8d77\u6765\u66f4\u65b9\u4fbf\u3002<\/p>\n
# \u514b\u9686Web UI\u9879\u76ee
\ncd ~\/ai-projects\/gemma-3
\ngit clone https:\/\/github.com\/example\/gemma-3-webui.git
\ncd gemma-3-webui<\/p>\n
# \u521b\u5efaPython\u865a\u62df\u73af\u5883
\npython3.11 -m venv venv
\nsource venv\/bin\/activate<\/p>\n
# \u5b89\u88c5\u4f9d\u8d56\u5305
\npip install –upgrade pip
\npip install torch torchvision torchaudio –index-url https:\/\/download.pytorch.org\/whl\/cpu
\npip install -r requirements.txt<\/p>\n
# \u914d\u7f6e\u6a21\u578b\u8def\u5f84
\ncat > config.yaml << EOF
\nmodel:
\n path: "\/root\/ai-models\/LLM-Research\/gemma-3-12b-it"
\n name: "gemma-3-12b-it"
\n device: "cpu" # \u4f7f\u7528CPU\u63a8\u7406<\/p>\n
server:
\n host: "0.0.0.0"
\n port: 7860
\n workers: 2<\/p>\n
generation:
\n max_tokens: 512
\n temperature: 0.7
\n top_p: 0.9
\nEOF<\/p>\n
\u572832GB\u5185\u5b58\u7684\u670d\u52a1\u5668\u4e0a\u8fd0\u884c23GB\u7684\u6a21\u578b，\u5185\u5b58\u7ba1\u7406\u81f3\u5173\u91cd\u8981\u3002\u4ee5\u4e0b\u662f\u5173\u952e\u7684\u4f18\u5316\u914d\u7f6e：<\/p>\n
# \u5728model_service.py\u4e2d\u6dfb\u52a0\u5185\u5b58\u4f18\u5316\u914d\u7f6e
\nimport os
\nos.environ["OMP_NUM_THREADS"] = "8" # \u9650\u5236OpenMP\u7ebf\u7a0b\u6570
\nos.environ["MKL_NUM_THREADS"] = "8" # \u9650\u5236MKL\u7ebf\u7a0b\u6570<\/p>\n
# \u4f7f\u7528\u5185\u5b58\u6620\u5c04\u52a0\u8f7d\u5927\u6a21\u578b
\ndef load_model_with_mmap(model_path):
\n """
\n \u4f7f\u7528\u5185\u5b58\u6620\u5c04\u65b9\u5f0f\u52a0\u8f7d\u6a21\u578b，\u51cf\u5c11\u5185\u5b58\u5360\u7528
\n """
\n from transformers import AutoModelForCausalLM, AutoTokenizer<\/p>\n
# \u5173\u952e\u914d\u7f6e：\u4f7f\u7528\u4f4e\u5185\u5b58\u6a21\u5f0f
\n model = AutoModelForCausalLM.from_pretrained(
\n model_path,
\n torch_dtype=torch.float16, # \u4f7f\u7528\u534a\u7cbe\u5ea6\u51cf\u5c11\u5185\u5b58
\n low_cpu_mem_usage=True, # \u4f4eCPU\u5185\u5b58\u4f7f\u7528\u6a21\u5f0f
\n device_map="cpu", # \u6307\u5b9a\u4f7f\u7528CPU
\n offload_folder="offload" # \u6ea2\u51fa\u65f6\u4e34\u65f6\u5b58\u50a8\u76ee\u5f55
\n )<\/p>\n
tokenizer = AutoTokenizer.from_pretrained(model_path)
\n return model, tokenizer<\/p>\n
\u914d\u7f6e\u5b8c\u6210\u540e，\u542f\u52a8\u670d\u52a1\u5e76\u9a8c\u8bc1\u662f\u5426\u6b63\u5e38\u8fd0\u884c：<\/p>\n
# \u542f\u52a8Web UI\u670d\u52a1
\ncd ~\/ai-projects\/gemma-3\/gemma-3-webui
\n.\/start.sh<\/p>\n
# \u67e5\u770b\u670d\u52a1\u72b6\u6001
\n.\/manage.sh status<\/p>\n
# \u76d1\u63a7\u5185\u5b58\u4f7f\u7528\u60c5\u51b5
\nwatch -n 1 "free -h && echo '—' && ps aux | grep python | grep -v grep"<\/p>\n
\u670d\u52a1\u542f\u52a8\u540e，\u5728\u6d4f\u89c8\u5668\u4e2d\u8bbf\u95ee http:\/\/\u670d\u52a1\u5668IP:7860，\u5e94\u8be5\u80fd\u770b\u5230Web UI\u754c\u9762\u3002\u9996\u6b21\u52a0\u8f7d\u6a21\u578b\u53ef\u80fd\u9700\u89811-2\u5206\u949f，\u8fd9\u662f\u6b63\u5e38\u73b0\u8c61\u3002<\/p>\n
\u4e3a\u4e86\u5168\u9762\u8bc4\u4f30\u6a21\u578b\u572832GB\u5185\u5b58\u73af\u5883\u4e0b\u7684\u7a33\u5b9a\u6027，\u6211\u8bbe\u8ba1\u4e86\u4ee5\u4e0b\u51e0\u4e2a\u6d4b\u8bd5\u7ef4\u5ea6：<\/p>\n
\u6838\u5fc3\u6d4b\u8bd5\u6307\u6807：<\/p>\n
\u6211\u51c6\u5907\u4e86\u56db\u7c7b\u6d4b\u8bd5\u7528\u4f8b，\u8986\u76d6\u4e0d\u540c\u7684\u4f7f\u7528\u573a\u666f：<\/p>\n
# \u6d4b\u8bd5\u7528\u4f8b\u5b9a\u4e49
\ntest_cases = {
\n "\u77ed\u6587\u672c\u5bf9\u8bdd": [
\n "\u4f60\u597d，\u8bf7\u4ecb\u7ecd\u4e00\u4e0b\u4f60\u81ea\u5df1",
\n "\u4eca\u5929\u7684\u5929\u6c14\u600e\u4e48\u6837？",
\n "Python\u662f\u4ec0\u4e48？",
\n "\u8c22\u8c22\u4f60\u7684\u5e2e\u52a9"
\n ],<\/p>\n
"\u4ee3\u7801\u751f\u6210\u4efb\u52a1": [
\n "\u5199\u4e00\u4e2aPython\u51fd\u6570\u8ba1\u7b97\u6590\u6ce2\u90a3\u5951\u6570\u5217",
\n "\u7528JavaScript\u5b9e\u73b0\u4e00\u4e2a\u7b80\u5355\u7684\u5f85\u529e\u4e8b\u9879\u5217\u8868",
\n "\u5199\u4e00\u4e2aSQL\u67e5\u8be2，\u627e\u51fa\u9500\u552e\u989d\u6700\u9ad8\u7684\u4ea7\u54c1",
\n "\u5b9e\u73b0\u4e00\u4e2a\u5feb\u901f\u6392\u5e8f\u7b97\u6cd5"
\n ],<\/p>\n
"\u957f\u6587\u672c\u751f\u6210": [
\n "\u5199\u4e00\u7bc7\u5173\u4e8e\u4eba\u5de5\u667a\u80fd\u672a\u6765\u53d1\u5c55\u7684\u77ed\u6587，\u7ea6300\u5b57",
\n "\u8be6\u7ec6\u89e3\u91ca\u4ec0\u4e48\u662f\u673a\u5668\u5b66\u4e60，\u5305\u62ec\u4e3b\u8981\u7c7b\u578b\u548c\u5e94\u7528\u573a\u666f",
\n "\u63cf\u8ff0\u4e00\u4e0b\u4e91\u8ba1\u7b97\u7684\u57fa\u672c\u6982\u5ff5\u548c\u4f18\u52bf"
\n ],<\/p>\n
"\u590d\u6742\u63a8\u7406\u4efb\u52a1": [
\n "\u5982\u679c\u6211\u67091000\u5143，\u5e74\u5229\u73875%，\u5b583\u5e74，\u6700\u540e\u80fd\u62ff\u5230\u591a\u5c11\u94b1？\u8bf7\u5199\u51fa\u8ba1\u7b97\u8fc7\u7a0b",
\n "\u6bd4\u8f83Python\u548cJava\u5728Web\u5f00\u53d1\u4e2d\u7684\u4f18\u7f3a\u70b9",
\n "\u89e3\u91ca\u4e00\u4e0bTCP\u4e09\u6b21\u63e1\u624b\u7684\u8fc7\u7a0b"
\n ]
\n}<\/p>\n
\u4e3a\u4e86\u7cfb\u7edf\u6027\u5730\u8fdb\u884c\u6d4b\u8bd5，\u6211\u7f16\u5199\u4e86\u4e00\u4e2a\u81ea\u52a8\u5316\u6d4b\u8bd5\u811a\u672c：<\/p>\n
import requests
\nimport time
\nimport psutil
\nimport json
\nfrom datetime import datetime<\/p>\n
class StabilityTester:
\n def __init__(self, base_url="http:\/\/localhost:7860"):
\n self.base_url = base_url
\n self.results = []<\/p>\n
def get_memory_usage(self):
\n """\u83b7\u53d6\u5f53\u524d\u5185\u5b58\u4f7f\u7528\u60c5\u51b5"""
\n memory = psutil.virtual_memory()
\n return {
\n "total": memory.total \/ (1024**3), # GB
\n "available": memory.available \/ (1024**3),
\n "used": memory.used \/ (1024**3),
\n "percent": memory.percent
\n }<\/p>\n
def send_request(self, prompt, max_tokens=512):
\n """\u53d1\u9001\u8bf7\u6c42\u5230\u6a21\u578b\u5e76\u8bb0\u5f55\u6027\u80fd\u6570\u636e"""
\n start_time = time.time()
\n memory_before = self.get_memory_usage()<\/p>\n
try:
\n response = requests.post(
\n f"{self.base_url}\/api\/generate",
\n json={
\n "prompt": prompt,
\n "max_tokens": max_tokens,
\n "temperature": 0.7
\n },
\n timeout=300 # 5\u5206\u949f\u8d85\u65f6
\n )<\/p>\n
elapsed_time = time.time() – start_time
\n memory_after = self.get_memory_usage()<\/p>\n
result = {
\n "timestamp": datetime.now().isoformat(),
\n "prompt": prompt[:50] + "…" if len(prompt) > 50 else prompt,
\n "response_time": elapsed_time,
\n "memory_before": memory_before,
\n "memory_after": memory_after,
\n "memory_increase": memory_after["used"] – memory_before["used"],
\n "success": response.status_code == 200,
\n "response_length": len(response.text) if response.status_code == 200 else 0
\n }<\/p>\n
return result<\/p>\n
except Exception as e:
\n elapsed_time = time.time() – start_time
\n return {
\n "timestamp": datetime.now().isoformat(),
\n "prompt": prompt[:50] + "…" if len(prompt) > 50 else prompt,
\n "response_time": elapsed_time,
\n "error": str(e),
\n "success": False
\n }<\/p>\n
def run_stress_test(self, test_cases, duration_hours=8):
\n """\u8fd0\u884c\u538b\u529b\u6d4b\u8bd5"""
\n print(f"\u5f00\u59cb\u538b\u529b\u6d4b\u8bd5，\u6301\u7eed\u65f6\u95f4：{duration_hours}\u5c0f\u65f6")
\n print(f"\u5f00\u59cb\u65f6\u95f4：{datetime.now()}")
\n print(f"\u521d\u59cb\u5185\u5b58\u4f7f\u7528：{self.get_memory_usage()}")<\/p>\n
end_time = time.time() + duration_hours * 3600
\n test_count = 0<\/p>\n
while time.time() < end_time:
\n # \u968f\u673a\u9009\u62e9\u6d4b\u8bd5\u7528\u4f8b\u7c7b\u578b
\n import random
\n case_type = random.choice(list(test_cases.keys()))
\n prompt = random.choice(test_cases[case_type])<\/p>\n
# \u53d1\u9001\u8bf7\u6c42
\n result = self.send_request(prompt)
\n self.results.append(result)
\n test_count += 1<\/p>\n
# \u6253\u5370\u8fdb\u5ea6
\n if test_count % 10 == 0:
\n print(f"\u5df2\u5b8c\u6210 {test_count} \u6b21\u6d4b\u8bd5，\u5f53\u524d\u5185\u5b58\u4f7f\u7528：{self.get_memory_usage()['percent']}%")<\/p>\n
# \u968f\u673a\u7b49\u5f85\u65f6\u95f4，\u6a21\u62df\u771f\u5b9e\u4f7f\u7528\u573a\u666f
\n time.sleep(random.uniform(1, 5))<\/p>\n
print(f"\u538b\u529b\u6d4b\u8bd5\u5b8c\u6210，\u603b\u5171\u6267\u884c {test_count} \u6b21\u6d4b\u8bd5")
\n return self.results<\/p>\n
def generate_report(self):
\n """\u751f\u6210\u6d4b\u8bd5\u62a5\u544a"""
\n successful_tests = [r for r in self.results if r.get("success", False)]
\n failed_tests = [r for r in self.results if not r.get("success", False)]<\/p>\n
report = {
\n "summary": {
\n "total_tests": len(self.results),
\n "successful_tests": len(successful_tests),
\n "failed_tests": len(failed_tests),
\n "success_rate": len(successful_tests) \/ len(self.results) * 100 if self.results else 0
\n },
\n "performance": {
\n "avg_response_time": sum(r["response_time"] for r in successful_tests) \/ len(successful_tests) if successful_tests else 0,
\n "max_response_time": max(r["response_time"] for r in successful_tests) if successful_tests else 0,
\n "min_response_time": min(r["response_time"] for r in successful_tests) if successful_tests else 0,
\n "avg_memory_increase": sum(r.get("memory_increase", 0) for r in successful_tests) \/ len(successful_tests) if successful_tests else 0
\n },
\n "memory_usage": {
\n "peak_usage": max(r.get("memory_after", {}).get("percent", 0) for r in self.results) if self.results else 0,
\n "avg_usage": sum(r.get("memory_after", {}).get("percent", 0) for r in self.results) \/ len(self.results) if self.results else 0
\n }
\n }<\/p>\n
return report<\/p>\n
# \u8fd0\u884c\u6d4b\u8bd5
\nif __name__ == "__main__":
\n tester = StabilityTester()
\n results = tester.run_stress_test(test_cases, duration_hours=8)
\n report = tester.generate_report()<\/p>\n
# \u4fdd\u5b58\u7ed3\u679c
\n with open("stability_test_report.json", "w") as f:
\n json.dump({"results": results, "report": report}, f, indent=2)<\/p>\n
print("\u6d4b\u8bd5\u62a5\u544a\u5df2\u4fdd\u5b58\u5230 stability_test_report.json")<\/p>\n
\u7ecf\u8fc78\u5c0f\u65f6\u7684\u8fde\u7eed\u538b\u529b\u6d4b\u8bd5，\u4ee5\u4e0b\u662f\u5185\u5b58\u4f7f\u7528\u7684\u5173\u952e\u6570\u636e：<\/p>\n
\u5185\u5b58\u4f7f\u7528\u7edf\u8ba1\u8868：<\/p>\n
| \u6a21\u578b\u52a0\u8f7d\u65f6<\/td>\n | 25.3 GB<\/td>\n | 26.8 GB<\/td>\n | \u00b11.5 GB<\/td>\n<\/tr>\n | ||||||||||||||||||||
| \u7a7a\u95f2\u72b6\u6001<\/td>\n | 24.1 GB<\/td>\n | 24.5 GB<\/td>\n | \u00b10.4 GB<\/td>\n<\/tr>\n | ||||||||||||||||||||
| \u63a8\u7406\u8fc7\u7a0b\u4e2d<\/td>\n | 25.7 GB<\/td>\n | 27.2 GB<\/td>\n | \u00b11.5 GB<\/td>\n<\/tr>\n | ||||||||||||||||||||
| \u538b\u529b\u6d4b\u8bd5\u671f<\/td>\n | 26.3 GB<\/td>\n | 28.1 GB<\/td>\n | \u00b11.8 GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u5173\u952e\u53d1\u73b0：<\/p>\n 5.2 \u54cd\u5e94\u6027\u80fd\u8868\u73b0<\/h4>\n\u54cd\u5e94\u65f6\u95f4\u7edf\u8ba1\u8868：<\/p>\n
|