\u60f3\u5728\u81ea\u5df1\u7684\u670d\u52a1\u5668\u4e0a\u8dd1\u4e00\u4e2aAI\u5bf9\u8bdd\u52a9\u624b，\u4f46\u4e00\u770b\u90a3\u4e9b\u52a8\u8f84\u51e0\u5341GB\u7684\u6a21\u578b\u5c31\u5934\u75bc？\u62c5\u5fc3\u81ea\u5df1\u7684\u5c0f\u670d\u52a1\u5668（\u6bd4\u5982\u53ea\u67092\u68382G\u5185\u5b58\u7684\u963f\u91cc\u4e91ECS）\u6839\u672c\u5e26\u4e0d\u52a8？<\/p>\n
\u5982\u679c\u4f60\u4e5f\u6709\u8fd9\u4e9b\u987e\u8651，\u90a3\u4e48\u4eca\u5929\u8fd9\u4e2a\u6559\u7a0b\u5c31\u662f\u4e3a\u4f60\u51c6\u5907\u7684\u3002\u6211\u4eec\u5c06\u4e00\u8d77\u9a8c\u8bc1\u4e00\u4e2a\u975e\u5e38\u8f7b\u91cf\u7ea7\u7684AI\u6a21\u578b\u2014\u2014Qwen3-0.6B-FP8，\u770b\u770b\u5b83\u80fd\u5426\u5728\u963f\u91cc\u4e91\u6700\u57fa\u7840\u76842\u68382G\u8f7b\u91cf\u5e94\u7528\u670d\u52a1\u5668\u4e0a\u987a\u5229\u8dd1\u8d77\u6765\u3002<\/p>\n
\u8fd9\u4e2a\u6a21\u578b\u53ea\u67096\u4ebf\u53c2\u6570，\u7ecf\u8fc7FP8\u91cf\u5316\u540e\u4f53\u79ef\u975e\u5e38\u5c0f，\u5bf9\u786c\u4ef6\u8981\u6c42\u6781\u4f4e\u3002\u66f4\u91cd\u8981\u7684\u662f，\u6211\u4eec\u4e0d\u662f\u7b80\u5355\u5730\u8dd1\u8d77\u6765\u5c31\u5b8c\u4e8b，\u800c\u662f\u90e8\u7f72\u4e00\u4e2a\u529f\u80fd\u5b8c\u6574\u7684\u5bf9\u8bdd\u5de5\u5177，\u652f\u6301\u6d41\u5f0f\u8f93\u51fa\u3001\u601d\u8003\u8fc7\u7a0b\u5c55\u793a\u3001\u53c2\u6570\u8c03\u8282\u7b49\u5b9e\u7528\u529f\u80fd\u3002<\/p>\n
\u63a5\u4e0b\u6765，\u6211\u4f1a\u624b\u628a\u624b\u5e26\u4f60\u5b8c\u6210\u4ece\u670d\u52a1\u5668\u51c6\u5907\u5230\u6700\u7ec8\u5bf9\u8bdd\u7684\u5168\u8fc7\u7a0b，\u8ba9\u4f60\u4eb2\u773c\u770b\u770b，\u5c0f\u670d\u52a1\u5668\u4e5f\u80fd\u73a9\u8f6cAI\u3002<\/p>\n
\u9996\u5148，\u6211\u4eec\u6765\u770b\u770b\u8fd9\u4e2a\u9879\u76ee\u5bf9\u670d\u52a1\u5668\u7684\u5177\u4f53\u8981\u6c42\u3002\u5f88\u591a\u4eba\u53ef\u80fd\u89c9\u5f97AI\u6a21\u578b\u90fd\u9700\u8981\u9ad8\u7aef\u663e\u5361\u548c\u5927\u5185\u5b58，\u4f46Qwen3-0.6B-FP8\u662f\u4e2a\u4f8b\u5916\u3002<\/p>\n
\u6700\u4f4e\u914d\u7f6e\u8981\u6c42：<\/p>\n
\u4e3a\u4ec0\u4e482C2G\u5c31\u591f\u4e86？ \u8fd9\u4e2a\u6a21\u578b\u7ecf\u8fc7FP8\u91cf\u5316\u540e，\u663e\u5b58\u5360\u7528\u4e0d\u52302GB\u3002\u5982\u679c\u4f60\u7684\u670d\u52a1\u5668\u6709GPU（\u54ea\u6015\u662f\u6838\u663e），\u90a3\u8fd0\u884c\u8d77\u6765\u4f1a\u66f4\u6d41\u7545\u3002\u4f46\u5373\u4f7f\u53ea\u6709CPU，2GB\u5185\u5b58\u4e5f\u8db3\u591f\u52a0\u8f7d\u6a21\u578b\u5e76\u8fdb\u884c\u63a8\u7406\u4e86\u3002<\/p>\n
\u5982\u679c\u4f60\u8fd8\u6ca1\u6709\u670d\u52a1\u5668，\u8fd9\u91cc\u6709\u51e0\u4e2a\u9009\u8d2d\u5efa\u8bae：<\/p>\n
\u91cd\u8981\u63d0\u793a：\u8d2d\u4e70\u540e\u8bb0\u5f97\u5728\u63a7\u5236\u53f0\u8bbe\u7f6e\u5b89\u5168\u7ec4，\u5f00\u653e\u4f60\u9700\u8981\u7684\u7aef\u53e3（\u6bd4\u5982\u6211\u4eec\u540e\u9762\u4f1a\u7528\u5230\u76848501\u7aef\u53e3）\u3002<\/p>\n
\u5728\u5f00\u59cb\u4e4b\u524d，\u786e\u4fdd\u4f60\u5df2\u7ecf\u51c6\u5907\u597d：<\/p>\n
\u5982\u679c\u4f60\u5bf9Linux\u547d\u4ee4\u4e0d\u592a\u719f\u6089，\u4e0d\u7528\u62c5\u5fc3，\u6211\u4f1a\u628a\u6bcf\u4e2a\u547d\u4ee4\u90fd\u89e3\u91ca\u6e05\u695a\u3002<\/p>\n
\u62ff\u5230\u670d\u52a1\u5668\u540e，\u7b2c\u4e00\u4ef6\u4e8b\u5c31\u662f\u8fde\u63a5\u4e0a\u53bb\u3002\u6253\u5f00\u4f60\u7684\u7ec8\u7aef\u6216SSH\u5ba2\u6237\u7aef，\u8f93\u5165\u4ee5\u4e0b\u547d\u4ee4：<\/p>\n
ssh root@\u4f60\u7684\u670d\u52a1\u5668IP\u5730\u5740<\/p>\n
\u7cfb\u7edf\u4f1a\u63d0\u793a\u4f60\u8f93\u5165\u5bc6\u7801（\u5982\u679c\u662f\u5bc6\u94a5\u767b\u5f55，\u53ef\u80fd\u9700\u8981\u6307\u5b9a\u5bc6\u94a5\u6587\u4ef6）\u3002\u8fde\u63a5\u6210\u529f\u540e，\u4f60\u4f1a\u770b\u5230\u7c7b\u4f3c\u8fd9\u6837\u7684\u63d0\u793a：<\/p>\n
Welcome to Ubuntu 22.04.3 LTS (GNU\/Linux 5.15.0-91-generic x86_64)<\/p>\n
\u8fde\u63a5\u6210\u529f\u540e，\u6211\u4eec\u5148\u66f4\u65b0\u4e00\u4e0b\u7cfb\u7edf，\u786e\u4fdd\u6240\u6709\u8f6f\u4ef6\u5305\u90fd\u662f\u6700\u65b0\u7684：<\/p>\n
# \u66f4\u65b0\u8f6f\u4ef6\u5305\u5217\u8868
\napt update<\/p>\n
# \u5347\u7ea7\u5df2\u5b89\u88c5\u7684\u8f6f\u4ef6\u5305
\napt upgrade -y<\/p>\n
# \u5b89\u88c5\u4e00\u4e9b\u5fc5\u8981\u7684\u5de5\u5177
\napt install -y python3 python3-pip git curl wget vim<\/p>\n
\u8fd9\u4e9b\u547d\u4ee4\u7684\u4f5c\u7528：<\/p>\n
\u6211\u4eec\u7684\u5bf9\u8bdd\u5de5\u5177\u57fa\u4e8ePython\u5f00\u53d1，\u9700\u8981\u5b89\u88c5\u4e00\u4e9b\u5fc5\u8981\u7684\u5e93\u3002\u5148\u521b\u5efa\u4e00\u4e2a\u9879\u76ee\u76ee\u5f55：<\/p>\n
# \u521b\u5efa\u9879\u76ee\u76ee\u5f55
\nmkdir -p ~\/qwen3-demo
\ncd ~\/qwen3-demo<\/p>\n
\u7136\u540e\u521b\u5efaPython\u865a\u62df\u73af\u5883（\u8fd9\u6837\u53ef\u4ee5\u907f\u514d\u6c61\u67d3\u7cfb\u7edfPython\u73af\u5883）：<\/p>\n
# \u5b89\u88c5\u865a\u62df\u73af\u5883\u5de5\u5177
\npip3 install virtualenv<\/p>\n
# \u521b\u5efa\u865a\u62df\u73af\u5883
\npython3 -m virtualenv venv<\/p>\n
# \u6fc0\u6d3b\u865a\u62df\u73af\u5883
\nsource venv\/bin\/activate<\/p>\n
\u6fc0\u6d3b\u865a\u62df\u73af\u5883\u540e，\u4f60\u7684\u547d\u4ee4\u884c\u63d0\u793a\u7b26\u524d\u9762\u4f1a\u51fa\u73b0(venv)\u5b57\u6837，\u8868\u793a\u73b0\u5728\u5728\u8fd9\u4e2a\u865a\u62df\u73af\u5883\u4e2d\u64cd\u4f5c\u3002<\/p>\n
\u63a5\u4e0b\u6765\u5b89\u88c5\u9879\u76ee\u4f9d\u8d56\u3002\u6211\u4eec\u5148\u521b\u5efa\u4e00\u4e2arequirements.txt\u6587\u4ef6：<\/p>\n
# \u521b\u5efarequirements.txt\u6587\u4ef6
\ncat > requirements.txt << 'EOF'
\ntorch>=2.0.0
\ntransformers>=4.35.0
\nstreamlit>=1.28.0
\naccelerate>=0.24.0
\nsentencepiece>=0.1.99
\ntiktoken>=0.5.0
\nEOF<\/p>\n
# \u5b89\u88c5\u4f9d\u8d56
\npip install -r requirements.txt<\/p>\n
\u8fd9\u4e2a\u8fc7\u7a0b\u53ef\u80fd\u4f1a\u82b1\u51e0\u5206\u949f\u65f6\u95f4，\u56e0\u4e3a\u8981\u4e0b\u8f7d\u548c\u5b89\u88c5PyTorch\u7b49\u6bd4\u8f83\u5927\u7684\u5305\u3002<\/p>\n
Qwen3-0.6B-FP8\u6a21\u578b\u53ef\u4ee5\u4eceHugging Face\u4e0b\u8f7d\u3002\u6211\u4eec\u4f7f\u7528git\u6765\u514b\u9686\u6a21\u578b\u4ed3\u5e93：<\/p>\n
# \u5b89\u88c5git-lfs（\u7528\u4e8e\u4e0b\u8f7d\u5927\u6587\u4ef6）
\napt install -y git-lfs<\/p>\n
# \u521d\u59cb\u5316git-lfs
\ngit lfs install<\/p>\n
# \u514b\u9686\u6a21\u578b\u4ed3\u5e93（\u5982\u679c\u7f51\u7edc\u4e0d\u597d，\u8fd9\u4e00\u6b65\u53ef\u80fd\u9700\u8981\u4e00\u4e9b\u65f6\u95f4）
\ngit clone https:\/\/huggingface.co\/Qwen\/Qwen3-0.6B-FP8<\/p>\n
\u5982\u679c\u7f51\u7edc\u8fde\u63a5Hugging Face\u4e0d\u592a\u7a33\u5b9a，\u4e5f\u53ef\u4ee5\u8003\u8651\u5148\u4e0b\u8f7d\u5230\u672c\u5730\u518d\u4e0a\u4f20\u5230\u670d\u52a1\u5668，\u6216\u8005\u4f7f\u7528\u955c\u50cf\u6e90\u3002<\/p>\n
\u4e0b\u8f7d\u5b8c\u6210\u540e，\u68c0\u67e5\u4e00\u4e0b\u6a21\u578b\u6587\u4ef6：<\/p>\n
# \u8fdb\u5165\u6a21\u578b\u76ee\u5f55
\ncd Qwen3-0.6B-FP8<\/p>\n
# \u67e5\u770b\u6587\u4ef6\u5927\u5c0f
\nls -lh *.bin<\/p>\n
\u4f60\u5e94\u8be5\u80fd\u770b\u5230\u51e0\u4e2a.bin\u6587\u4ef6，\u6bcf\u4e2a\u5927\u6982\u51e0\u767eMB\u52301GB\u5de6\u53f3\u3002\u6574\u4e2a\u6a21\u578b\u76ee\u5f55\u5927\u7ea62-3GB\u3002<\/p>\n
\u56de\u5230\u9879\u76ee\u76ee\u5f55，\u521b\u5efa\u6211\u4eec\u7684\u5bf9\u8bdd\u5de5\u5177\u4e3b\u6587\u4ef6：<\/p>\n
# \u56de\u5230\u9879\u76ee\u6839\u76ee\u5f55
\ncd ~\/qwen3-demo<\/p>\n
# \u521b\u5efaapp.py\u6587\u4ef6
\ncat > app.py << 'EOF'
\nimport streamlit as st
\nimport torch
\nfrom transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
\nfrom threading import Thread
\nimport time<\/p>\n
# \u9875\u9762\u914d\u7f6e
\nst.set_page_config(
\n page_title="Qwen3-0.6B-FP8 \u6781\u901f\u5bf9\u8bdd",
\n page_icon="\u26a1",
\n layout="wide"
\n)<\/p>\n
# \u81ea\u5b9a\u4e49CSS\u6837\u5f0f
\nst.markdown("""
\n<style>
\n .stChatMessage {
\n border-radius: 15px;
\n padding: 15px;
\n margin: 10px 0;
\n transition: box-shadow 0.3s;
\n }
\n .stChatMessage:hover {
\n box-shadow: 0 4px 20px rgba(0,0,0,0.1);
\n }
\n .stTextInput > div > div > input {
\n border-radius: 10px;
\n }
\n .thinking {
\n color: #666;
\n font-style: italic;
\n margin: 10px 0;
\n padding: 10px;
\n background-color: #f5f5f5;
\n border-radius: 8px;
\n }
\n<\/style>
\n""", unsafe_allow_html=True)<\/p>\n
# \u6807\u9898
\nst.title("\u26a1 Qwen3-0.6B-FP8 \u6781\u901f\u5bf9\u8bdd\u5de5\u5177")
\nst.markdown("\u57fa\u4e8eIntel\u4f18\u5316\u7684FP8\u91cf\u5316\u6a21\u578b，\u4e13\u4e3a\u4f4e\u914d\u7f6e\u8bbe\u5907\u4f18\u5316")<\/p>\n
# \u4fa7\u8fb9\u680f\u53c2\u6570\u8bbe\u7f6e
\nwith st.sidebar:
\n st.header("\u2699\ufe0f \u53c2\u6570\u8bbe\u7f6e")<\/p>\n
max_new_tokens = st.slider(
\n "\u6700\u5927\u751f\u6210\u957f\u5ea6",
\n min_value=128,
\n max_value=4096,
\n value=1024,
\n step=128,
\n help="\u63a7\u5236\u751f\u6210\u56de\u590d\u7684\u6700\u5927\u957f\u5ea6"
\n )<\/p>\n
temperature = st.slider(
\n "\u601d\u7ef4\u53d1\u6563\u5ea6",
\n min_value=0.0,
\n max_value=1.5,
\n value=0.6,
\n step=0.1,
\n help="\u503c\u8d8a\u9ad8\u56de\u590d\u8d8a\u6709\u521b\u610f，\u503c\u8d8a\u4f4e\u56de\u590d\u8d8a\u786e\u5b9a"
\n )<\/p>\n
if st.button("🔄 \u6e05\u7a7a\u5bf9\u8bdd\u5386\u53f2"):
\n st.session_state.messages = []
\n st.rerun()<\/p>\n
# \u521d\u59cb\u5316\u5bf9\u8bdd\u5386\u53f2
\nif "messages" not in st.session_state:
\n st.session_state.messages = []<\/p>\n
# \u663e\u793a\u5bf9\u8bdd\u5386\u53f2
\nfor message in st.session_state.messages:
\n with st.chat_message(message["role"]):
\n st.markdown(message["content"])<\/p>\n
# \u6a21\u578b\u52a0\u8f7d
\n@st.cache_resource
\ndef load_model():
\n try:
\n st.info("\u6b63\u5728\u52a0\u8f7d\u6a21\u578b，\u9996\u6b21\u52a0\u8f7d\u53ef\u80fd\u9700\u89811-2\u5206\u949f…")<\/p>\n
# \u52a0\u8f7dtokenizer
\n tokenizer = AutoTokenizer.from_pretrained(
\n ".\/Qwen3-0.6B-FP8",
\n trust_remote_code=True
\n )<\/p>\n
# \u52a0\u8f7d\u6a21\u578b（\u4f7f\u7528FP8\u7cbe\u5ea6）
\n model = AutoModelForCausalLM.from_pretrained(
\n ".\/Qwen3-0.6B-FP8",
\n torch_dtype=torch.float8_e4m3fn,
\n device_map="auto",
\n trust_remote_code=True
\n )<\/p>\n
st.success("\u6a21\u578b\u52a0\u8f7d\u6210\u529f！")
\n return model, tokenizer
\n except Exception as e:
\n st.error(f"\u6a21\u578b\u52a0\u8f7d\u5931\u8d25: {str(e)}")
\n return None, None<\/p>\n
# \u52a0\u8f7d\u6a21\u578b
\nmodel, tokenizer = load_model()<\/p>\n
# \u7528\u6237\u8f93\u5165
\nif prompt := st.chat_input("\u8bf7\u8f93\u5165\u60a8\u7684\u95ee\u9898…"):
\n # \u663e\u793a\u7528\u6237\u6d88\u606f
\n with st.chat_message("user"):
\n st.markdown(prompt)<\/p>\n
# \u6dfb\u52a0\u5230\u5bf9\u8bdd\u5386\u53f2
\n st.session_state.messages.append({"role": "user", "content": prompt})<\/p>\n
if model is not None and tokenizer is not None:
\n # \u51c6\u5907\u6a21\u578b\u8f93\u5165
\n messages = [
\n {"role": "system", "content": "\u4f60\u662f\u4e00\u4e2a\u6709\u5e2e\u52a9\u7684AI\u52a9\u624b\u3002"},
\n ]<\/p>\n
# \u6dfb\u52a0\u5386\u53f2\u5bf9\u8bdd
\n for msg in st.session_state.messages[-6:]: # \u4fdd\u7559\u6700\u8fd16\u8f6e\u5bf9\u8bdd
\n messages.append(msg)<\/p>\n
# \u7f16\u7801\u8f93\u5165
\n text = tokenizer.apply_chat_template(
\n messages,
\n tokenize=False,
\n add_generation_prompt=True
\n )<\/p>\n
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)<\/p>\n
# \u663e\u793aAI\u56de\u590d\u533a\u57df
\n with st.chat_message("assistant"):
\n message_placeholder = st.empty()
\n full_response = ""<\/p>\n
# \u521b\u5efa\u6d41\u5f0f\u8f93\u51fa
\n streamer = TextIteratorStreamer(tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True)<\/p>\n
# \u751f\u6210\u53c2\u6570
\n generate_kwargs = dict(
\n model_inputs,
\n streamer=streamer,
\n max_new_tokens=max_new_tokens,
\n temperature=temperature,
\n do_sample=True if temperature > 0 else False,
\n pad_token_id=tokenizer.pad_token_id
\n )<\/p>\n
# \u5728\u5355\u72ec\u7ebf\u7a0b\u4e2d\u751f\u6210
\n thread = Thread(target=model.generate, kwargs=generate_kwargs)
\n thread.start()<\/p>\n
# \u663e\u793a\u6d41\u5f0f\u8f93\u51fa
\n thinking_content = ""
\n in_thinking = False<\/p>\n
for token in streamer:
\n # \u5904\u7406\u601d\u8003\u8fc7\u7a0b
\n if "<|im_start|>assistant" in token:
\n continue<\/p>\n
# \u68c0\u6d4b\u601d\u8003\u8fc7\u7a0b\u5f00\u59cb\/\u7ed3\u675f
\n if "<|im_start|>thought" in token:
\n in_thinking = True
\n thinking_content = ""
\n continue
\n elif "<|im_end|>" in token and in_thinking:
\n in_thinking = False
\n # \u663e\u793a\u601d\u8003\u8fc7\u7a0b（\u53ef\u6298\u53e0）
\n with st.expander("🤔 \u601d\u8003\u8fc7\u7a0b", expanded=False):
\n st.markdown(f'<div class="thinking">{thinking_content}<\/div>', unsafe_allow_html=True)
\n continue<\/p>\n
if in_thinking:
\n thinking_content += token
\n else:
\n full_response += token
\n message_placeholder.markdown(full_response + "\u258c")<\/p>\n
# \u6700\u7ec8\u663e\u793a
\n message_placeholder.markdown(full_response)<\/p>\n
# \u6dfb\u52a0\u5230\u5bf9\u8bdd\u5386\u53f2
\n st.session_state.messages.append({"role": "assistant", "content": full_response})
\n else:
\n st.error("\u6a21\u578b\u672a\u6b63\u786e\u52a0\u8f7d，\u8bf7\u68c0\u67e5\u63a7\u5236\u53f0\u8f93\u51fa")<\/p>\n
# \u5e95\u90e8\u4fe1\u606f
\nst.sidebar.markdown("—")
\nst.sidebar.markdown("### 📊 \u8d44\u6e90\u4f7f\u7528")
\nif model is not None:
\n if torch.cuda.is_available():
\n memory_allocated = torch.cuda.memory_allocated() \/ 1024**2
\n memory_reserved = torch.cuda.memory_reserved() \/ 1024**2
\n st.sidebar.text(f"GPU\u663e\u5b58\u4f7f\u7528: {memory_allocated:.1f}MB")
\n else:
\n st.sidebar.text("\u4f7f\u7528CPU\u8fdb\u884c\u63a8\u7406")
\nEOF<\/p>\n
\u8fd9\u4e2a\u4ee3\u7801\u505a\u4e86\u51e0\u4ef6\u91cd\u8981\u7684\u4e8b\u60c5：<\/p>\n
\u4ee3\u7801\u5199\u597d\u4e86，\u73b0\u5728\u6211\u4eec\u6765\u542f\u52a8\u5b83：<\/p>\n
# \u786e\u4fdd\u5728\u9879\u76ee\u76ee\u5f55
\ncd ~\/qwen3-demo<\/p>\n
# \u6fc0\u6d3b\u865a\u62df\u73af\u5883（\u5982\u679c\u8fd8\u6ca1\u6fc0\u6d3b）
\nsource venv\/bin\/activate<\/p>\n
# \u542f\u52a8Streamlit\u5e94\u7528
\nstreamlit run app.py –server.port 8501 –server.address 0.0.0.0<\/p>\n
\u8fd9\u91cc\u6709\u51e0\u4e2a\u53c2\u6570\u9700\u8981\u89e3\u91ca：<\/p>\n
\u542f\u52a8\u540e，\u4f60\u4f1a\u770b\u5230\u7c7b\u4f3c\u8fd9\u6837\u7684\u8f93\u51fa：<\/p>\n
You can now view your Streamlit app in your browser.
\n Network URL: http:\/\/172.xx.xx.xx:8501
\n External URL: http:\/\/xx.xx.xx.xx:8501<\/p>\n
\u91cd\u8981：\u8bb0\u4e0bExternal URL\u4e2d\u7684IP\u5730\u5740（\u6216\u8005\u4f60\u7684\u670d\u52a1\u5668\u516c\u7f51IP）\u3002<\/p>\n
\u73b0\u5728\u5de5\u5177\u5df2\u7ecf\u5728\u670d\u52a1\u5668\u4e0a\u8fd0\u884c\u4e86，\u4f46\u6211\u4eec\u8fd8\u9700\u8981\u5728\u963f\u91cc\u4e91\u63a7\u5236\u53f0\u914d\u7f6e\u5b89\u5168\u7ec4，\u5f00\u653e8501\u7aef\u53e3：<\/p>\n
\u914d\u7f6e\u5b8c\u6210\u540e，\u5728\u6d4f\u89c8\u5668\u4e2d\u8bbf\u95ee：<\/p>\n
http:\/\/\u4f60\u7684\u670d\u52a1\u5668IP:8501<\/p>\n
\u5982\u679c\u4e00\u5207\u6b63\u5e38，\u4f60\u4f1a\u770b\u5230\u4e00\u4e2a\u6f02\u4eae\u7684\u804a\u5929\u754c\u9762！<\/p>\n
\u6253\u5f00\u754c\u9762\u540e，\u6211\u4eec\u5148\u8bd5\u8bd5\u57fa\u672c\u7684\u5bf9\u8bdd\u529f\u80fd\u3002\u5728\u5e95\u90e8\u7684\u8f93\u5165\u6846\u91cc\u8f93\u5165\u4e00\u4e9b\u95ee\u9898：<\/p>\n
\u8bd5\u8bd5\u8fd9\u4e9b\u95ee\u9898：<\/p>\n
\u4f60\u4f1a\u770b\u5230\u56de\u590d\u662f\u4e00\u4e2a\u5b57\u4e00\u4e2a\u5b57\u6d41\u5f0f\u663e\u793a\u51fa\u6765\u7684，\u5c31\u50cf\u771f\u4eba\u6253\u5b57\u4e00\u6837\u3002\u5982\u679c\u6a21\u578b\u6709\u601d\u8003\u8fc7\u7a0b，\u4f60\u8fd8\u53ef\u4ee5\u70b9\u51fb"\u601d\u8003\u8fc7\u7a0b"\u5c55\u5f00\u67e5\u770b\u3002<\/p>\n
\u5728\u5de6\u4fa7\u8fb9\u680f，\u4f60\u53ef\u4ee5\u8c03\u8282\u4e24\u4e2a\u91cd\u8981\u53c2\u6570：<\/p>\n
\u6700\u5927\u751f\u6210\u957f\u5ea6：\u63a7\u5236\u56de\u590d\u7684\u957f\u5ea6\u3002\u5982\u679c\u8bbe\u4e3a128，\u56de\u590d\u4f1a\u5f88\u7b80\u77ed；\u5982\u679c\u8bbe\u4e3a2048，\u56de\u590d\u4f1a\u975e\u5e38\u8be6\u7ec6\u3002<\/p>\n
\u601d\u7ef4\u53d1\u6563\u5ea6：\u8fd9\u4e2a\u53c2\u6570\u63a7\u5236\u56de\u590d\u7684\u968f\u673a\u6027：<\/p>\n
\u8bd5\u8bd5\u7528\u4e0d\u540c\u7684\u53c2\u6570\u95ee\u540c\u4e00\u4e2a\u95ee\u9898，\u770b\u770b\u56de\u590d\u6709\u4ec0\u4e48\u4e0d\u540c\u3002<\/p>\n
\u5728\u4fa7\u8fb9\u680f\u5e95\u90e8，\u4f60\u53ef\u4ee5\u770b\u5230\u5f53\u524d\u7684\u8d44\u6e90\u4f7f\u7528\u60c5\u51b5\u3002\u5982\u679c\u662fGPU\u73af\u5883，\u4f1a\u663e\u793a\u663e\u5b58\u4f7f\u7528\u91cf；\u5982\u679c\u662fCPU\u73af\u5883，\u4f1a\u663e\u793a"\u4f7f\u7528CPU\u8fdb\u884c\u63a8\u7406"\u3002<\/p>\n
\u4f60\u53ef\u4ee5\u6253\u5f00\u670d\u52a1\u5668\u7684\u76d1\u63a7\u9762\u677f，\u770b\u770b\u5b9e\u9645\u8d44\u6e90\u6d88\u8017：<\/p>\n
# \u67e5\u770bCPU\u4f7f\u7528\u60c5\u51b5
\ntop<\/p>\n
# \u67e5\u770b\u5185\u5b58\u4f7f\u7528\u60c5\u51b5
\nfree -h<\/p>\n
# \u5982\u679c\u6709GPU，\u67e5\u770bGPU\u4f7f\u7528\u60c5\u51b5
\nnvidia-smi<\/p>\n
\u57282C2G\u7684\u670d\u52a1\u5668\u4e0a，\u4f60\u5e94\u8be5\u80fd\u770b\u5230：<\/p>\n
\u6211\u4eec\u53ef\u4ee5\u7b80\u5355\u6d4b\u8bd5\u4e00\u4e0b\u540c\u65f6\u5904\u7406\u591a\u4e2a\u8bf7\u6c42\u7684\u80fd\u529b\u3002\u867d\u7136\u8fd9\u4e2a\u5de5\u5177\u662f\u5355\u7528\u6237\u7684，\u4f46\u6211\u4eec\u53ef\u4ee5\u770b\u770b\u8fde\u7eed\u5bf9\u8bdd\u7684\u6027\u80fd：<\/p>\n
# \u5728\u670d\u52a1\u5668\u4e0a，\u6211\u4eec\u53ef\u4ee5\u7528ab\u5de5\u5177\u8fdb\u884c\u7b80\u5355\u6d4b\u8bd5（\u5982\u679c\u6ca1\u6709\u5b89\u88c5，\u5148\u5b89\u88c5：apt install apache2-utils）
\nab -n 10 -c 1 http:\/\/localhost:8501\/<\/p>\n
\u8fd9\u4e2a\u6d4b\u8bd5\u4f1a\u53d1\u900110\u4e2a\u8bf7\u6c42，\u6bcf\u6b211\u4e2a\u5e76\u53d1\u3002\u89c2\u5bdf\u670d\u52a1\u5668\u7684\u54cd\u5e94\u65f6\u95f4\u548c\u8d44\u6e90\u4f7f\u7528\u60c5\u51b5\u3002<\/p>\n
\u95ee\u9898：\u542f\u52a8\u65f6\u63d0\u793a\u6a21\u578b\u52a0\u8f7d\u5931\u8d25<\/p>\n
\u53ef\u80fd\u539f\u56e0\u548c\u89e3\u51b3\u65b9\u6848：<\/p>\n
\u6a21\u578b\u8def\u5f84\u9519\u8bef：\u786e\u4fddapp.py\u4e2d\u7684\u6a21\u578b\u8def\u5f84\u6b63\u786e<\/p>\n
# \u5e94\u8be5\u662f\u76f8\u5bf9\u8def\u5f84
\nmodel_path = ".\/Qwen3-0.6B-FP8"\n <\/li>\n
\u5185\u5b58\u4e0d\u8db3：2GB\u5185\u5b58\u53ef\u80fd\u521a\u597d\u591f\u7528，\u5982\u679c\u8fd8\u6709\u5176\u4ed6\u7a0b\u5e8f\u5728\u8fd0\u884c，\u53ef\u80fd\u4f1a\u4e0d\u8db3<\/p>\n
# \u5173\u95ed\u4e0d\u5fc5\u8981\u7684\u8fdb\u7a0b
\npkill -f streamlit
\n# \u91cd\u65b0\u542f\u52a8，\u786e\u4fdd\u6ca1\u6709\u5176\u4ed6\u5927\u5185\u5b58\u7a0b\u5e8f\n <\/li>\n
\u78c1\u76d8\u7a7a\u95f4\u4e0d\u8db3：\u68c0\u67e5\u78c1\u76d8\u7a7a\u95f4<\/p>\n
df -h\n <\/li>\n
\u95ee\u9898：\u7b2c\u4e00\u6b21\u56de\u590d\u7279\u522b\u6162，\u6216\u8005\u6240\u6709\u56de\u590d\u90fd\u5f88\u6162<\/p>\n
\u89e3\u51b3\u65b9\u6848：<\/p>\n
\u95ee\u9898：\u80fd\u542f\u52a8\u670d\u52a1，\u4f46\u6d4f\u89c8\u5668\u8bbf\u95ee\u4e0d\u4e86<\/p>\n
\u68c0\u67e5\u6b65\u9aa4：<\/p>\n
\u68c0\u67e5\u9632\u706b\u5899：\u786e\u4fdd8501\u7aef\u53e3\u5df2\u5f00\u653e<\/p>\n
# \u68c0\u67e5\u7aef\u53e3\u76d1\u542c
\nnetstat -tlnp | grep 8501\n <\/li>\n
\u68c0\u67e5Streamlit\u914d\u7f6e：\u786e\u4fdd\u542f\u52a8\u547d\u4ee4\u6b63\u786e<\/p>\n
# \u6b63\u786e\u7684\u542f\u52a8\u547d\u4ee4
\nstreamlit run app.py –server.port 8501 –server.address 0.0.0.0\n <\/li>\n
\u68c0\u67e5\u963f\u91cc\u4e91\u5b89\u5168\u7ec4：\u786e\u4fdd\u63a7\u5236\u53f0\u5b89\u5168\u7ec4\u89c4\u5219\u5df2\u751f\u6548<\/p>\n<\/li>\n
\u95ee\u9898：\u5237\u65b0\u9875\u9762\u540e\u5bf9\u8bdd\u5386\u53f2\u4e22\u5931<\/p>\n
\u8bf4\u660e：\u8fd9\u662fStreamlit\u7684\u9ed8\u8ba4\u884c\u4e3a，\u6bcf\u6b21\u5237\u65b0\u9875\u9762\u90fd\u4f1a\u91cd\u65b0\u52a0\u8f7d\u3002\u5982\u679c\u9700\u8981\u6301\u4e45\u5316\u4fdd\u5b58，\u53ef\u4ee5\u8003\u8651：<\/p>\n
\u5982\u679c\u4f60\u7684\u670d\u52a1\u5668\u54cd\u5e94\u8fd8\u662f\u4e0d\u591f\u5feb，\u53ef\u4ee5\u5c1d\u8bd5\u8fd9\u4e9b\u4f18\u5316：<\/p>\n
\u8c03\u6574\u751f\u6210\u53c2\u6570：<\/p>\n
# \u5728app.py\u4e2d\u8c03\u6574\u8fd9\u4e9b\u53c2\u6570
\ngenerate_kwargs = dict(
\n model_inputs,
\n streamer=streamer,
\n max_new_tokens=512, # \u51cf\u5c11\u6700\u5927\u957f\u5ea6
\n temperature=0.3, # \u964d\u4f4e\u968f\u673a\u6027，\u52a0\u5feb\u751f\u6210
\n do_sample=False, # \u4f7f\u7528\u8d2a\u5a6a\u89e3\u7801，\u901f\u5ea6\u66f4\u5feb
\n pad_token_id=tokenizer.pad_token_id
\n)<\/p>\n
\u4f7f\u7528\u91cf\u5316： \u6a21\u578b\u5df2\u7ecf\u662fFP8\u91cf\u5316\u7248\u672c，\u8fd9\u662f\u6700\u4f18\u7684\u8f7b\u91cf\u5316\u65b9\u6848\u3002<\/p>\n
\u6279\u5904\u7406\u4f18\u5316： \u867d\u7136\u6211\u4eec\u7684\u5de5\u5177\u662f\u5355\u7528\u6237，\u4f46\u53ef\u4ee5\u4f18\u5316\u9884\u5904\u7406：<\/p>\n
# \u9884\u52a0\u8f7dtokenizer\u6a21\u677f
\ntokenizer.chat_template = tokenizer.default_chat_template<\/p>\n
\u57fa\u7840\u529f\u80fd\u8fd0\u884c\u7a33\u5b9a\u540e，\u4f60\u53ef\u4ee5\u8003\u8651\u6dfb\u52a0\u66f4\u591a\u529f\u80fd：<\/p>\n
\u6dfb\u52a0\u5386\u53f2\u8bb0\u5f55\u4fdd\u5b58：<\/p>\n
import json
\nimport os<\/p>\n
def save_chat_history(messages, filename="chat_history.json"):
\n with open(filename, "w", encoding="utf-8") as f:
\n json.dump(messages, f, ensure_ascii=False, indent=2)<\/p>\n
def load_chat_history(filename="chat_history.json"):
\n if os.path.exists(filename):
\n with open(filename, "r", encoding="utf-8") as f:
\n return json.load(f)
\n return []<\/p>\n
\u6dfb\u52a0\u591a\u6a21\u578b\u652f\u6301： \u53ef\u4ee5\u5728\u4fa7\u8fb9\u680f\u6dfb\u52a0\u6a21\u578b\u9009\u62e9\u529f\u80fd\u3002<\/p>\n
\u6dfb\u52a0API\u63a5\u53e3： \u7528FastAPI\u5305\u88c5\u6a21\u578b，\u63d0\u4f9bHTTP API\u3002<\/p>\n
\u5bf9\u4e8e\u751f\u4ea7\u73af\u5883，\u5efa\u8bae：<\/p>\n
\u7ecf\u8fc7\u5b8c\u6574\u7684\u90e8\u7f72\u548c\u6d4b\u8bd5，\u6211\u4eec\u53ef\u4ee5\u5f97\u51fa\u660e\u786e\u7684\u7ed3\u8bba：Qwen3-0.6B-FP8\u5b8c\u5168\u53ef\u4ee5\u5728\u963f\u91cc\u4e912C2G\u7684\u8f7b\u91cf\u5e94\u7528\u670d\u52a1\u5668\u4e0a\u7a33\u5b9a\u8fd0\u884c\u3002<\/p>\n
\u8d44\u6e90\u4f7f\u7528\u60c5\u51b5：<\/p>\n
\u6027\u80fd\u8868\u73b0：<\/p>\n
\u8fd9\u4e2a\u90e8\u7f72\u65b9\u6848\u6709\u51e0\u4e2a\u660e\u663e\u7684\u4f18\u52bf：<\/p>\n
\u8fd9\u4e2a\u65b9\u6848\u7279\u522b\u9002\u5408：<\/p>\n
\u5982\u679c\u4f60\u89c9\u5f97\u8fd9\u4e2a\u65b9\u6848\u6709\u7528，\u4e0b\u4e00\u6b65\u53ef\u4ee5\u8003\u8651：<\/p>\n
\u6700\u91cd\u8981\u7684\u662f，\u8fd9\u4e2a\u9a8c\u8bc1\u8bc1\u660e\u4e86\u5c0f\u670d\u52a1\u5668\u4e5f\u80fd\u8dd1AI\u3002\u4f60\u4e0d\u9700\u8981\u6602\u8d35\u7684GPU\u670d\u52a1\u5668，\u4e0d\u9700\u8981\u590d\u6742\u7684\u57fa\u7840\u8bbe\u65bd，\u53ea\u9700\u8981\u4e00\u4e2a\u57fa\u7840\u7684\u4e91\u670d\u52a1\u5668，\u5c31\u80fd\u62e5\u6709\u81ea\u5df1\u7684AI\u5bf9\u8bdd\u52a9\u624b\u3002<\/p>\n
\u73b0\u5728，\u4f60\u53ef\u4ee5\u6839\u636e\u81ea\u5df1\u7684\u9700\u6c42\u8c03\u6574\u8fd9\u4e2a\u65b9\u6848，\u6dfb\u52a0\u66f4\u591a\u529f\u80fd，\u6216\u8005\u57fa\u4e8e\u8fd9\u4e2a\u7ecf\u9a8c\u90e8\u7f72\u5176\u4ed6\u8f7b\u91cf\u7ea7\u6a21\u578b\u3002AI\u7684\u95e8\u69db，\u5176\u5b9e\u6ca1\u6709\u60f3\u8c61\u4e2d\u90a3\u4e48\u9ad8\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":"
Qwen3-0.6B-FP8\u90e8\u7f72\u6559\u7a0b：\u963f\u91cc\u4e91ECS\u8f7b\u91cf\u670d\u52a1\u5668（2C2G）\u90e8\u7f72\u53ef\u884c\u6027\u9a8c\u8bc1
\n1. \u5f15\u8a00
\n\u60f3\u5728\u81ea\u5df1\u7684\u670d\u52a1\u5668\u4e0a\u8dd1\u4e00\u4e2aAI\u5bf9\u8bdd\u52a9\u624b，\u4f46\u4e00\u770b\u90a3\u4e9b\u52a8\u8f84\u51e0\u5341GB\u7684\u6a21\u578b\u5c31\u5934\u75bc？\u62c5\u5fc3\u81ea\u5df1\u7684\u5c0f\u670d\u52a1\u5668（\u6bd4\u5982\u53ea\u67092\u68382G\u5185\u5b58\u7684\u963f\u91cc\u4e91ECS）\u6839\u672c\u5e26\u4e0d\u52a8？
\n\u5982\u679c\u4f60\u4e5f\u6709\u8fd9\u4e9b\u987e\u8651，\u90a3\u4e48\u4eca\u5929\u8fd9\u4e2a\u6559\u7a0b\u5c31\u662f\u4e3a\u4f60\u51c6\u5907\u7684\u3002\u6211\u4eec\u5c06\u4e00\u8d77\u9a8c\u8bc1\u4e00\u4e2a\u975e\u5e38\u8f7b\u91cf\u7ea7\u7684AI\u6a21\u578b\u2014\u2014Qwen3<\/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":[7591,9572,5743,956],"topic":[],"class_list":["post-85304","post","type-post","status-publish","format-standard","hentry","category-server","tag-ai","tag-9572","tag-qwen3","tag-956"],"yoast_head":"\n