refactor(xtask): 整理多模型服务的方案，支持通过配置文件配置模型

Signed-off-by: YdrMaster <ydrml@hotmail.com>

Author: YdrMaster

Cargo.lock

@@ -89,14 +89,26 @@ cargo service --help
 ```plaintext
 web service
 
-Usage: xtask service [OPTIONS] --port <PORT> <MODEL>
+Usage: xtask service [OPTIONS] --port <PORT> <FILE>
 
 Arguments:
-  <MODEL>
+  <FILE>
 
 Options:
+  -p, --port <PORT>
+      --no-cuda-graph
+      --name <NAME>
       --gpus <GPUS>
       --max-steps <MAX_STEPS>
-  -p, --port <PORT>
+      --think
   -h, --help
 ```
+
+通过 TOML 配置文件可以配置多模型服务。示例格式：
+
+```toml
+[model-name]
+path = "model-path"
+think = true
+max-steps = 2048
+```

README.md

@@ -17,6 +17,7 @@ ratatui = "0.29"
 
 serde.workspace = true
 serde_json = "1.0"
+toml = "0.8"
 tokio = { version = "1.45", features = ["rt-multi-thread", "net"] }
 hyper = { version = "1.6", features = ["http1", "server"] }
 hyper-util = { version = "0.1", features = ["http1", "tokio", "server"] }

xtask/Cargo.toml

@@ -55,23 +55,27 @@ struct BaseArgs {
 
 impl BaseArgs {
     fn gpus(&self) -> Box<[c_int]> {
-        self.gpus
-            .as_ref()
-            .map(|devices| {
-                static NUM_REGEX: LazyLock<Regex> = LazyLock::new(|| Regex::new(r"\d+").unwrap());
-                NUM_REGEX
-                    .find_iter(devices)
-                    .map(|c| c.as_str().parse().unwrap())
-                    .collect()
-            })
-            .unwrap_or_else(|| [0].into())
+        parse_gpus(self.gpus.as_deref())
     }
 
     fn max_steps(&self) -> usize {
         self.max_steps.unwrap_or(1000)
     }
 }
 
+fn parse_gpus(config: Option<&str>) -> Box<[c_int]> {
+    config
+        .as_ref()
+        .map(|devices| {
+            static NUM_REGEX: LazyLock<Regex> = LazyLock::new(|| Regex::new(r"\d+").unwrap());
+            NUM_REGEX
+                .find_iter(devices)
+                .map(|c| c.as_str().parse().unwrap())
+                .collect()
+        })
+        .unwrap_or_else(|| [0].into())
+}
+
 mod macros {
     macro_rules! print_now {
         ($($arg:tt)*) => {{

xtask/src/main.rs

@@ -1,6 +1,8 @@
-use super::*;
+use super::V1_CHAT_COMPLETIONS;
 use log::{info, trace, warn};
-use openai_struct::CreateChatCompletionStreamResponse;
+use openai_struct::{
+    ChatCompletionRequestMessage, CreateChatCompletionRequest, CreateChatCompletionStreamResponse,
+};
 use reqwest::header::{CONTENT_TYPE, HeaderMap, HeaderValue};
 use std::{env::VarError, time::Instant};
 use tokio::time::Duration;