Merge remote-tracking branch 'origin/main'
This commit is contained in:
liushuang
commit
7976f4dd3a
@ -1,5 +1,3 @@
|
||||
# 常用的 Docker 镜像脚本和命令
|
||||
|
||||
## 安装 DPanel
|
||||
|
||||
```shell
|
||||
@ -1,6 +1,4 @@
|
||||
# RTX4090笔电操作记录
|
||||
|
||||
```shell
|
||||
```
|
||||
# 因清华大学开源镜像站 HTTP/403 换了中科大的镜像站，配置信息存放在这里
|
||||
cat /etc/apt/sources.list
|
||||
|
||||
@ -34,24 +32,14 @@ pip install vllm -i http://mirrors.cloud.tencent.com/pypi/simple --extra-index-u
|
||||
# 安装 modelscope
|
||||
pip install modelscope -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host mirrors.aliyun.com
|
||||
|
||||
# 拉取 gpt-oss-20b 模型
|
||||
# 拉取 gpt-oss-20b 模型，由于显存不足，运行失败了
|
||||
modelscope download --model openai-mirror/gpt-oss-20b --local_dir /home/ss/vllm-py12/gpt-oss-20b
|
||||
|
||||
# 运行 gpt-oss-20b 模型失败，移动端的 RTX4090 只有 16GB 显存，至少需要 16~24GB 显存
|
||||
vllm serve \
|
||||
/home/ss/vllm-py12/gpt-oss-20b \
|
||||
--port 18777 \
|
||||
--api-key token_lcfc \
|
||||
--served-model-name gpt-oss-20b \
|
||||
--gpu-memory-utilization 0.95 \
|
||||
--tool-call-parser openai \
|
||||
--enable-auto-tool-choice
|
||||
|
||||
# Qwen3-8b 也需要 16~24GB显存，所以下载了 Qwen3-0.6B
|
||||
# 下载了 Qwen3-0.6B
|
||||
modelscope download --model Qwen/Qwen3-0.6B --local_dir /home/ss/vllm-py12/qwen3-06b
|
||||
|
||||
# 运行 Qwen3-8b
|
||||
vllm serve /home/ss/vllm-py12/qwen3-06b \
|
||||
# 运行 Qwen3-0.6B
|
||||
nohup vllm serve /home/ss/vllm-py12/qwen3-06b \
|
||||
--host 0.0.0.0 \
|
||||
--port 8000 \
|
||||
--served-model-name Qwen3-0.6B \
|
||||
@ -59,6 +47,24 @@ vllm serve /home/ss/vllm-py12/qwen3-06b \
|
||||
--dtype auto \
|
||||
--gpu-memory-utilization 0.9 \
|
||||
--max-model-len 32768 \
|
||||
--trust-remote-code
|
||||
--trust-remote-code \
|
||||
>> /home/ss/vllm-py12/vllm.log 2>&1 \
|
||||
& echo $! > /home/ss/vllm-py12/vllm.pid
|
||||
|
||||
# 安装了抓包工具 tshark 和 ngrep
|
||||
sudo apt install ngrep
|
||||
sudo apt-get install tshark
|
||||
|
||||
# 通过 java 脚本调用 tshark 提取关键日志
|
||||
sudo nohup bash /home/ss/vllm-py12/tshark_bash.sh >> /home/ss/vllm-py12/tshark_bash.log 2>&1 & echo $! > /home/ss/vllm-py12/tshark_bash.pid
|
||||
|
||||
# 运行了1个定时任务脚本，清理 tshark 的临时文件并重启 java 脚本
|
||||
sudo nohup /home/ss/vllm-py12/timer_bash.sh > /home/ss/vllm-py12/timer_bash.log 2>&1 & echo $! > /home/ss/vllm-py12/timer_bash.pid
|
||||
|
||||
# 杀死上面2个进程的命令
|
||||
sudo kill -9 $(cat /home/ss/vllm-py12/timer_bash.pid)
|
||||
sudo kill -9 $(cat /home/ss/vllm-py12/tshark_bash.pid)
|
||||
|
||||
# 清理日志
|
||||
cd /home/ss/vllm-py12 && rm -rf timer_bash.log tshark_bash.log shark.log
|
||||
```
|
||||
145
202510_使用Java执行tshark命令抓包.md
Normal file
145
202510_使用Java执行tshark命令抓包.md
Normal file
@ -0,0 +1,145 @@
|
||||
## 脚本说明
|
||||
|
||||
- 在 Linux 运行`TShark.java`必须先安装**tshark**，在 Windows 上运行必须先安装**Wireshark**客户端，在安装时勾选**tshark**组件。
|
||||
- 必须使用超级管理员账号运行`TShark.java`，这是**tshark**命令的限制，在 Windows 上可以使用管理员权限打开命令行运行。
|
||||
- Java 脚本基于**Java21**开发，可直接使用`java TShark.java`运行，运行后会在当前目录下创建一个**shark.log**文件，保存抓包日志。
|
||||
- 建议创建一个 bash 或者 cmd 脚本，在 bash 或 cmd 脚本中配置好 JDK 环境变量，然后再使用管理员权限运行 bash 或 cmd 脚本。
|
||||
|
||||
##### TShark.java
|
||||
|
||||
```java
|
||||
import java.io.*;
|
||||
import java.nio.file.Path;
|
||||
import java.time.ZoneId;
|
||||
import java.util.List;
|
||||
import java.util.TimeZone;
|
||||
|
||||
public final class TShark {
|
||||
|
||||
public static void main(String[] args) throws Exception {
|
||||
TimeZone.setDefault(TimeZone.getTimeZone(ZoneId.of("Asia/Shanghai")));
|
||||
Path path = Path.of("shark.log");
|
||||
File sharkLog = path.toFile();
|
||||
if (!sharkLog.exists() && !sharkLog.createNewFile()) {
|
||||
throw new RuntimeException("create shark.log failure");
|
||||
}
|
||||
// DateTimeFormatter timeFormatter = DateTimeFormatter.ofPattern("yyyy-MM-dd HH:mm:ss");
|
||||
Process process = startSharkProcess();
|
||||
System.out.println("start tshark success");
|
||||
startProcessSharkThread(process, sharkLog);
|
||||
System.out.println("start process thread success");
|
||||
int exitCode = process.waitFor();
|
||||
System.out.println("tshark process is dead, exit code: " + exitCode);
|
||||
}
|
||||
|
||||
private static void startProcessSharkThread(Process process, File sharkLog) {
|
||||
Thread thread = new Thread(() -> {
|
||||
try (BufferedReader reader = new BufferedReader(new InputStreamReader(process.getInputStream()));
|
||||
BufferedWriter fileWriter = new BufferedWriter(new FileWriter(sharkLog, true))) {
|
||||
String line;
|
||||
while ((line = reader.readLine()) != null) {
|
||||
fileWriter.write(line + System.lineSeparator());
|
||||
}
|
||||
fileWriter.flush();
|
||||
} catch (Exception e) {
|
||||
e.printStackTrace();
|
||||
}
|
||||
});
|
||||
thread.setName("TShark-thread");
|
||||
thread.start();
|
||||
}
|
||||
|
||||
public static Process startSharkProcess() throws IOException {
|
||||
List<String> cmd = List.of(
|
||||
"tshark",
|
||||
"-l",
|
||||
"-i",
|
||||
"\\Device\\NPF_{807C63AC-179D-4AC8-BD56-85CE8AA179DB}",
|
||||
"-Y",
|
||||
"tcp.port == 33000 && http.request.method == \"POST\"",
|
||||
"-V"
|
||||
);
|
||||
ProcessBuilder processBuilder = new ProcessBuilder(cmd);
|
||||
processBuilder.redirectErrorStream(true);
|
||||
return processBuilder.start();
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
##### 通过 NEWAPI 调用 vLLM 运行的 gpt-oss-120b 模型
|
||||
|
||||
```java
|
||||
import java.net.URI;
|
||||
import java.net.http.HttpClient;
|
||||
import java.net.http.HttpRequest;
|
||||
import java.net.http.HttpResponse;
|
||||
import java.time.Duration;
|
||||
import java.util.concurrent.CompletableFuture;
|
||||
|
||||
public class HttpPost {
|
||||
|
||||
public static void main(String[] args) {
|
||||
try (HttpClient httpClient = HttpClient.newHttpClient()) {
|
||||
HttpRequest httpRequest = HttpRequest.newBuilder()
|
||||
.uri(URI.create(apiUrl))
|
||||
.header("Content-Type", "application/json")
|
||||
.header("Authorization", apiKey)
|
||||
.timeout(Duration.ofMinutes(5L))
|
||||
.POST(HttpRequest.BodyPublishers.ofString(request))
|
||||
.build();
|
||||
CompletableFuture<Void> completableFuture = httpClient.sendAsync(httpRequest, HttpResponse.BodyHandlers.ofLines()).thenAccept(response -> {
|
||||
System.out.println("SSE success status " + response.statusCode() + ", response body: ");
|
||||
response.body().forEach(System.out::println);
|
||||
}).exceptionally(throwable -> {
|
||||
System.out.println("SSE failure error message: " + throwable.getMessage());
|
||||
throwable.printStackTrace();
|
||||
return null;
|
||||
});
|
||||
try {
|
||||
completableFuture.get();
|
||||
} catch (Exception e) {
|
||||
e.printStackTrace();
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
static final String apiUrl = "http://10.159.252.49:33000/v1/chat/completions";
|
||||
|
||||
static final String apiKey = "Bearer sk-rbWEGdsaZ47e2hQFZt2xRHqWRZaipYlkyLqHdU2z9FlWj7D3";
|
||||
|
||||
static final String request = """
|
||||
{
|
||||
"model": "gpt-oss-120b",
|
||||
"messages": [
|
||||
{
|
||||
"role": "user",
|
||||
"content": "成龙今年71岁，请把这句话的信息，按照`user_response`结构生成一个JSON"
|
||||
}
|
||||
],
|
||||
"stream": true,
|
||||
"extra_body":{
|
||||
"structured_outputs": true,
|
||||
"guided_decoding": true
|
||||
},
|
||||
"response_format": {
|
||||
"type": "json_schema",
|
||||
"json_schema": {
|
||||
"name": "user_response",
|
||||
"schema": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"name": {"type": "string"},
|
||||
"age": {"type": "number"}
|
||||
},
|
||||
"required": ["name", "age"],
|
||||
"additionalProperties": false
|
||||
},
|
||||
"strict": true
|
||||
}
|
||||
}
|
||||
}
|
||||
""".strip();
|
||||
|
||||
}
|
||||
|
||||
```
|
||||
@ -1,5 +1,3 @@
|
||||
# 多智能体开发技术方案对比
|
||||
|
||||
## LLM基础应用框架评测对象
|
||||
|
||||
- Spring AI Alibaba（SAA）
|
||||
141
202510_定时监控vLLM进程.md
Normal file
141
202510_定时监控vLLM进程.md
Normal file
@ -0,0 +1,141 @@
|
||||
## 修改vLLM运行gpt-oss-120b的脚本
|
||||
|
||||
在 nohup 的启动命令末尾追加如下片段，保存 vLLM 的进程 PID 号。
|
||||
|
||||
```
|
||||
echo $! > /hook/gpt-oss-120b.pid
|
||||
```
|
||||
|
||||
## 创建一个定时任务 bash 脚本
|
||||
|
||||
```
|
||||
vim /hook/timer_bash.sh
|
||||
```
|
||||
|
||||
```
|
||||
#!/bin/bash
|
||||
|
||||
# 目标进程的 PID 文件路径
|
||||
TARGET_PID_FILE="/hook/gpt-oss-120b.pid"
|
||||
|
||||
# tshark_bash 的 PID 文件路径
|
||||
TSARK_PID_FILE="/hook/tshark_bash.pid"
|
||||
|
||||
# tshark_bash 脚本路径
|
||||
TSARK_SCRIPT_PATH="/hook/tshark_bash.sh"
|
||||
|
||||
# tshark_bash 日志文件路径
|
||||
TSARK_LOG_FILE="/hook/tshark_bash.log"
|
||||
|
||||
# shark.log 文件路径
|
||||
SHARK_LOG_FILE="/hook/shark.log"
|
||||
|
||||
# 临时目录下要删除的 pcapng 文件路径模式
|
||||
PCAPNG_FILES_PATTERN="/tmp/*.pcapng"
|
||||
|
||||
# 5分钟检查一次 vllm 进程是否存活
|
||||
# 若 vllm 进程存活，清理 tshark 的临时文件，然后重启 tshark 脚本继续抓包
|
||||
# 若 vllm 进程已死亡则 kill tshark 的进程
|
||||
|
||||
while true; do
|
||||
if [[ -f "$TARGET_PID_FILE" ]]; then
|
||||
TARGET_PID=$(cat "$TARGET_PID_FILE" 2>/dev/null)
|
||||
if [[ "$TARGET_PID" =~ ^[0-9]+$ ]] && ps -p "$TARGET_PID" > /dev/null 2>&1; then
|
||||
echo "$(date): 目标进程 (PID: $TARGET_PID) 存活，执行相应操作..."
|
||||
if [[ -f "$TSARK_PID_FILE" ]]; then
|
||||
TSHARK_PID=$(cat "$TSARK_PID_FILE" 2>/dev/null)
|
||||
if [[ "$TSHARK_PID" =~ ^[0-9]+$ ]]; then
|
||||
kill -9 "$TSHARK_PID" 2>/dev/null
|
||||
echo "已强制杀死 tshark_bash 进程 (PID: $TSHARK_PID)"
|
||||
else
|
||||
echo "tshark_bash.pid 文件内容无效或为空，跳过 kill 操作"
|
||||
fi
|
||||
else
|
||||
echo "tshark_bash.pid 文件不存在，跳过 kill 操作"
|
||||
fi
|
||||
|
||||
if [[ -f "$SHARK_LOG_FILE" ]]; then
|
||||
rm -rf "$SHARK_LOG_FILE"
|
||||
echo "已删除 $SHARK_LOG_FILE"
|
||||
else
|
||||
echo "$SHARK_LOG_FILE 文件不存在，跳过删除"
|
||||
fi
|
||||
|
||||
if ls $PCAPNG_FILES_PATTERN >/dev/null 2>&1; then
|
||||
rm -rf $PCAPNG_FILES_PATTERN
|
||||
echo "已删除 $PCAPNG_FILES_PATTERN 文件"
|
||||
else
|
||||
echo "没有找到 $PCAPNG_FILES_PATTERN 文件，跳过删除"
|
||||
fi
|
||||
|
||||
echo "重启 tshark_bash.sh..."
|
||||
sudo nohup bash "$TSARK_SCRIPT_PATH" >> "$TSARK_LOG_FILE" 2>&1 &
|
||||
NEW_TSHARK_PID=$!
|
||||
echo "$NEW_TSHARK_PID" > "$TSARK_PID_FILE"
|
||||
echo "已启动 tshark_bash.sh，新 PID: $NEW_TSHARK_PID，已写入 $TSARK_PID_FILE"
|
||||
|
||||
else
|
||||
echo "$(date): 目标进程 (PID: ${TARGET_PID:-未知}) 不存在或已挂掉，执行清理操作..."
|
||||
|
||||
if [[ -f "$TSARK_PID_FILE" ]]; then
|
||||
TSHARK_PID=$(cat "$TSARK_PID_FILE" 2>/dev/null)
|
||||
if [[ "$TSHARK_PID" =~ ^[0-9]+$ ]]; then
|
||||
kill -9 "$TSHARK_PID" 2>/dev/null
|
||||
echo "已强制杀死 tshark_bash 进程 (PID: $TSHARK_PID)"
|
||||
else
|
||||
echo "tshark_bash.pid 文件内容无效或为空，跳过 kill 操作"
|
||||
fi
|
||||
else
|
||||
echo "tshark_bash.pid 文件不存在，跳过 kill 操作"
|
||||
fi
|
||||
|
||||
if ls $PCAPNG_FILES_PATTERN >/dev/null 2>&1; then
|
||||
rm -rf $PCAPNG_FILES_PATTERN
|
||||
echo "已删除 $PCAPNG_FILES_PATTERN 文件"
|
||||
else
|
||||
echo "没有找到 $PCAPNG_FILES_PATTERN 文件，跳过删除"
|
||||
fi
|
||||
|
||||
echo "目标进程已挂掉，已执行清理操作。"
|
||||
fi
|
||||
else
|
||||
echo "$(date): 目标 PID 文件 $TARGET_PID_FILE 不存在，跳过本次检查。"
|
||||
fi
|
||||
|
||||
sleep 300
|
||||
done
|
||||
```
|
||||
|
||||
## 定时任务
|
||||
|
||||
##### 第一次运行先赋权
|
||||
|
||||
```
|
||||
chmod +x vim /hook/timer_bash.sh
|
||||
```
|
||||
|
||||
##### 运行
|
||||
|
||||
```
|
||||
sudo nohup /hook/timer_bash.sh > /hook/timer_bash.log 2>&1 & echo $! > /hook/timer_bash.pid
|
||||
```
|
||||
|
||||
##### 停止
|
||||
|
||||
```
|
||||
kill $(cat /hook/timer_bash.pid)
|
||||
```
|
||||
|
||||
## 抓包脚本
|
||||
|
||||
##### 运行
|
||||
|
||||
```
|
||||
sudo nohup bash /hook/tshark_bash.sh >> /hook/tshark_bash.log 2>&1 & echo $! > /hook/tshark_bash.pid
|
||||
```
|
||||
|
||||
##### 停止
|
||||
|
||||
```
|
||||
kill -9 $(cat /hook/tshark_bash.pid)
|
||||
```
|
||||
@ -1,293 +1,20 @@
|
||||
# vllm 崩溃问题排查
|
||||
|
||||
## 问题原因
|
||||
|
||||
vllm在处理请求时，调用了 get_structured_output_key 函数，在处理 structured_outputs 参数时，由于不满足所有的情况，导致了抛出**No valid structured output parameter found**异常，该异常未被 EngineCore 捕获导致引擎崩溃，APIServer 发现引擎死了，自己也退出了进程。
|
||||
|
||||
## 使用抓包工具 tshark 监控 http 请求
|
||||
|
||||
### 在合适的目录下执行下面的命令，tshark 需要 root 权限，请使用 root 账号执行
|
||||
|
||||
##### 安装 tshark（已执行完毕）
|
||||
```
|
||||
sudo apt-get install tshark
|
||||
```
|
||||
|
||||
##### 创建一个 shell 脚本
|
||||
```
|
||||
vim hook_vllm_gpt-oss-120b.sh
|
||||
```
|
||||
脚本内容如下：
|
||||
```
|
||||
#!/bin/bash
|
||||
|
||||
# ========================================
|
||||
# 监控本地 v1/chat/completions 接口的 HTTP 请求与响应
|
||||
# 仅保留最近 20 条日志（含流式响应，如 SSE）
|
||||
# 使用 tshark + TCP 流跟踪
|
||||
# ========================================
|
||||
|
||||
# 配置
|
||||
INTERFACE="lo" # 本地回环接口
|
||||
PORT="8077"
|
||||
ENDPOINT="/v1/chat/completions"
|
||||
LOG_FILE="/hook/chat_completions.log"
|
||||
TEMP_LOG="/hook/chat_completions.tmp"
|
||||
PID_FILE="/hook/hook_vllm_gpt-oss-120b.pid"
|
||||
|
||||
# 检查是否已运行
|
||||
if [ -f "$PID_FILE" ]; then
|
||||
if ps -p $(cat "$PID_FILE") > /dev/null 2>&1; then
|
||||
echo "【错误】监控脚本已在运行 (PID: $(cat $PID_FILE))"
|
||||
exit 1
|
||||
else
|
||||
rm -f "$PID_FILE"
|
||||
fi
|
||||
fi
|
||||
|
||||
# 记录 PID
|
||||
echo $$ > "$PID_FILE"
|
||||
|
||||
# 清空日志
|
||||
> "$LOG_FILE"
|
||||
|
||||
echo "✅ 开始监控 http://127.0.0.1:$PORT$ENDPOINT"
|
||||
echo "📝 日志将保存到: $LOG_FILE"
|
||||
echo "⏳ 仅保留最近 20 条，按 Ctrl+C 停止"
|
||||
|
||||
# 信号处理：清理 PID 文件
|
||||
trap 'rm -f "$PID_FILE"; echo "⏹️ 监控已停止"; exit 0' SIGINT SIGTERM
|
||||
|
||||
# 使用 tshark 跟踪 TCP 流
|
||||
sudo tshark \
|
||||
-i "$INTERFACE" \
|
||||
-f "tcp port $PORT and host 127.0.0.1" \
|
||||
-q \
|
||||
-z "follow,tcp,ascii" \
|
||||
2>/dev/null | \
|
||||
stdbuf -oL awk -v endpoint="$ENDPOINT" -v log="$LOG_FILE" -v temp="$TEMP" '
|
||||
BEGIN {
|
||||
RS = "\n\n"
|
||||
ORS = ""
|
||||
in_request = 0
|
||||
buffer = ""
|
||||
count = 0
|
||||
}
|
||||
|
||||
# 分割流，识别每条 TCP 流
|
||||
{
|
||||
if (match($0, /GET|POST|PUT|DELETE|HTTP/) && index($0, endpoint)) {
|
||||
# 提取时间戳（tshark 输出第一行包含时间）
|
||||
if (match($0, /Following.*on port [0-9]+$/)) {
|
||||
ts = substr($0, RSTART, RLENGTH)
|
||||
gsub(/.*on/, "on", ts)
|
||||
} else {
|
||||
ts = "unknown time"
|
||||
}
|
||||
|
||||
# 提取请求行和头
|
||||
split($0, lines, /\n/)
|
||||
for (i=1; i<=length(lines); i++) {
|
||||
if (lines[i] ~ /(GET|POST|PUT|DELETE)/ && index(lines[i], endpoint)) {
|
||||
request_line = lines[i]
|
||||
}
|
||||
if (lines[i] ~ /Content-Type:/ || lines[i] ~ /Authorization:/ || lines[i] ~ /User-Agent:/) {
|
||||
headers = headers "\n " lines[i]
|
||||
}
|
||||
}
|
||||
|
||||
# 提取请求体（通常在空行后）
|
||||
body = ""
|
||||
in_body = 0
|
||||
for (i=1; i<=length(lines); i++) {
|
||||
if (lines[i] == "" || lines[i] ~ /HTTP\/[0-9.]/) {
|
||||
in_body = 1
|
||||
continue
|
||||
}
|
||||
if (in_body && lines[i] !~ /(No response found|Following)/) {
|
||||
body = body lines[i] "\n"
|
||||
}
|
||||
}
|
||||
|
||||
# 提取响应部分（HTTP/ 开头）
|
||||
response = ""
|
||||
for (i=1; i<=length(lines); i++) {
|
||||
if (lines[i] ~ /^HTTP\// && i > 1) {
|
||||
for (j=i; j<=length(lines); j++) {
|
||||
if (lines[j] !~ /Following/) {
|
||||
response = response lines[j] "\n"
|
||||
}
|
||||
}
|
||||
break
|
||||
}
|
||||
}
|
||||
|
||||
# 构造日志条目
|
||||
entry = "========================================\n"
|
||||
entry = entry "🕒 " ts "\n"
|
||||
entry = entry "📤 请求: " request_line "\n"
|
||||
if (headers != "") {
|
||||
entry = entry "📎 头部:\n" headers "\n"
|
||||
}
|
||||
if (body != "") {
|
||||
entry = entry "📦 请求体:\n" body "\n"
|
||||
}
|
||||
if (response != "") {
|
||||
entry = entry "📥 响应:\n" response "\n"
|
||||
}
|
||||
entry = entry "========================================\n\n"
|
||||
|
||||
# 写入日志并保留最近 20 条
|
||||
cmd = "echo -e \"" entry "\" >> " log " && tail -n 200 " log " > " temp " && mv " temp " " log
|
||||
system(cmd)
|
||||
|
||||
# 重置
|
||||
headers = ""
|
||||
body = ""
|
||||
response = ""
|
||||
}
|
||||
}
|
||||
'
|
||||
|
||||
# 正常退出时清理
|
||||
rm -f "$PID_FILE"
|
||||
```
|
||||
|
||||
##### 赋予执行权限
|
||||
```
|
||||
chmod +x hook_vllm_gpt-oss-120b.sh
|
||||
```
|
||||
|
||||
##### 后台运行脚本
|
||||
```
|
||||
nohup /hook/hook_vllm_gpt-oss-120b.sh > /dev/null 2>&1 &
|
||||
```
|
||||
|
||||
##### 查看请求日志
|
||||
|
||||
```
|
||||
tail -f /hook/chat_completions.log
|
||||
```
|
||||
|
||||
##### 停止脚本
|
||||
|
||||
```
|
||||
pkill -f hook_vllm_gpt-oss-120b.sh
|
||||
```
|
||||
|
||||
### 函数 get_structured_output_key 实现
|
||||
|
||||
#### 分支：release/v0.11.0
|
||||
由于外部调用 vllm 的 OpenAI API 服务时候，传入的请求参数让 vllm 调用了**get_structured_output_key**函数。在该函数里不能被正确处理，抛出了**No valid structured output parameter found**异常，该异常导致了 vllm 的 EngineCore 和 APIServer 进程死亡。
|
||||
|
||||
## 源码出处
|
||||
https://github.com/vllm-project/vllm/blob/releases/v0.11.0/vllm/v1/structured_output/request.py
|
||||
|
||||
```python
|
||||
def get_structured_output_key(sampling_params: SamplingParams) -> StructuredOutputKey:
|
||||
params = sampling_params.structured_outputs
|
||||
assert params is not None, "params can't be None."
|
||||
if params.json is not None:
|
||||
if not isinstance(params.json, str):
|
||||
json_str = json.dumps(params.json)
|
||||
else:
|
||||
json_str = params.json
|
||||
return (StructuredOutputOptions.JSON, json_str)
|
||||
elif params.json_object:
|
||||
return (StructuredOutputOptions.JSON_OBJECT, "")
|
||||
elif params.regex is not None:
|
||||
return (StructuredOutputOptions.REGEX, params.regex)
|
||||
elif params.choice is not None:
|
||||
if not isinstance(params.choice, str):
|
||||
json_str = json.dumps(params.choice)
|
||||
else:
|
||||
json_str = params.choice
|
||||
return (StructuredOutputOptions.CHOICE, json_str)
|
||||
elif params.grammar is not None:
|
||||
return (StructuredOutputOptions.GRAMMAR, params.grammar)
|
||||
elif params.structural_tag is not None:
|
||||
return (StructuredOutputOptions.STRUCTURAL_TAG, params.structural_tag)
|
||||
else:
|
||||
raise ValueError("No valid structured output parameter found")
|
||||
```
|
||||
|
||||
#### 分支：release/v0.10.2
|
||||
## 问题追踪
|
||||
|
||||
https://github.com/vllm-project/vllm/blob/releases/v0.10.2/vllm/v1/structured_output/request.py
|
||||
由于 vllm 没有提供 http 请求参数的日志打印，也没有集成监控 http 请求的三方工具。所以在 Ubuntu 上安装了 tshark 抓包工具。通过 Java 脚本启动 tshark 命令，并将 tshark 抓包到的日志内容写入到磁盘文件。下一次 vllm 崩溃时，根据磁盘文件存储的日志内容分析是什么参数导致了**get_structured_output_key**的异常。
|
||||
|
||||
```python
|
||||
def get_structured_output_key(
|
||||
sampling_params: SamplingParams) -> StructuredOutputKey:
|
||||
params = sampling_params.guided_decoding
|
||||
assert params is not None, "params can't be None."
|
||||
if params.json is not None:
|
||||
if not isinstance(params.json, str):
|
||||
json_str = json.dumps(params.json)
|
||||
else:
|
||||
json_str = params.json
|
||||
return (StructuredOutputOptions.JSON, json_str)
|
||||
elif params.json_object:
|
||||
return (StructuredOutputOptions.JSON_OBJECT, "")
|
||||
elif params.regex is not None:
|
||||
return (StructuredOutputOptions.REGEX, params.regex)
|
||||
elif params.choice is not None:
|
||||
if not isinstance(params.choice, str):
|
||||
json_str = json.dumps(params.choice)
|
||||
else:
|
||||
json_str = params.choice
|
||||
return (StructuredOutputOptions.CHOICE, json_str)
|
||||
elif params.grammar is not None:
|
||||
return (StructuredOutputOptions.GRAMMAR, params.grammar)
|
||||
elif params.structural_tag is not None:
|
||||
return (StructuredOutputOptions.STRUCTURAL_TAG, params.structural_tag)
|
||||
else:
|
||||
raise ValueError("No valid structured output parameter found")
|
||||
```
|
||||
> 2个版本的唯一区别，仅仅是提取`SamplingParams`的属性不一样，其它判断逻辑都是一致的
|
||||
> 使用 Java 执行 tshark 是为了抹掉 Linux 和 Windows 的平台差异，不用修改代码和命令即可直接运行。
|
||||
|
||||
### SamplingParams
|
||||
> Java 脚本内容：
|
||||
|
||||
https://github.com/vllm-project/vllm/blob/releases/v0.11.0/vllm/sampling_params.py
|
||||
|
||||
```python
|
||||
# 其它参数已省略
|
||||
class SamplingParams(
|
||||
💡初始化逻辑是没问题的，默认值是None
|
||||
structured_outputs: Optional[StructuredOutputsParams] = None
|
||||
guided_decoding: Optional[GuidedDecodingParams] = None
|
||||
|
||||
@staticmethod
|
||||
def from_optional(
|
||||
💡默认值也是None
|
||||
structured_outputs: Optional[StructuredOutputsParams] = None,
|
||||
) -> "SamplingParams":
|
||||
if guided_decoding is not None:
|
||||
warnings.warn(
|
||||
"guided_decoding is deprecated. This will be removed in "
|
||||
💡官方将在 v0.12.0 废弃 guided_decoding 参数，使用 structured_outputs 参数替代，在 v0.11.0 版本做了兼容，
|
||||
"v0.12.0 or v1.0.0, which ever is soonest. Please use "
|
||||
"structured_outputs instead.",
|
||||
DeprecationWarning,
|
||||
stacklevel=2)
|
||||
structured_outputs = guided_decoding
|
||||
guided_decoding = None
|
||||
return SamplingParams(
|
||||
structured_outputs=structured_outputs,
|
||||
)
|
||||
```
|
||||
|
||||
### StructuredOutputOptions
|
||||
|
||||
```python
|
||||
class StructuredOutputOptions(enum.Enum):
|
||||
JSON = enum.auto()
|
||||
JSON_OBJECT = enum.auto()
|
||||
REGEX = enum.auto()
|
||||
GRAMMAR = enum.auto()
|
||||
CHOICE = enum.auto()
|
||||
STRUCTURAL_TAG = enum.auto()
|
||||
```
|
||||
> 💡只支持这6种类型，每个类型都对应 structured_outputs 下面的一个不同的参数。
|
||||
|
||||
## 崩溃日志片段
|
||||
## 崩溃日志
|
||||
|
||||
```text
|
||||
[1;36m(EngineCore_DP0 pid=2738693)[0;0m ERROR 10-10 10:43:10 [core.py:710] EngineCore encountered a fatal error.
|
||||
Loading…
Reference in New Issue
Block a user