feat(inference): 添加流式输出功能，支持实时解码文本

examples/minimal-inference/src/main.rs

@@ -25,7 +25,7 @@ fn main() -> anyhow::Result<()> {
 
     println!("模型加载完成，耗时: {:.2?}", start.elapsed());
 
-    let prompt = "你知道今天的天气不";
+    let prompt = "我怕人家一说要改前端文件，明天咔嚓甩给我了";
     println!("\n用户: {}", prompt);
     println!("Assistant: ");
 
@@ -36,10 +36,18 @@ fn main() -> anyhow::Result<()> {
     };
 
     let start = Instant::now();
-    let response = model.generate(prompt, false, &config)?;
+    let _response = model.generate_stream(
+        prompt,
+        false,
+        &config,
+        |token| {
+            print!("{}", token);
+            std::io::Write::flush(&mut std::io::stdout()).ok();
+        },
+    )?;
     let elapsed = start.elapsed();
 
-    println!("{}", response);
+    println!();
     println!("\n生成耗时: {:.2?}", elapsed);
 
     Ok(())