feat(inference): 添加流式输出功能，支持实时解码文本

.workbuddy/memory/2026-07-01.md

@@ -14,3 +14,8 @@
 - 三个 crate 版本升级到 0.1.1，convert/inference 的 minicpm-core 依赖也更新到 0.1.1
 - 发布顺序：`cargo publish --allow-dirty --registry gitea -p minicpm-core` → `minicpm-convert` → `minicpm-inference`
 - 全部发布成功
+
+## 添加流式输出
+
+- `minicpm-inference/src/lib.rs`: 新增 `generate_stream` 函数和 `MiniCPM::generate_stream` 方法，每 token 通过回调 `impl FnMut(&str)` 输出解码文本
+- `examples/minimal-inference/src/main.rs`: 改为流式调用，`print!` + `flush` 实时输出

crates/minicpm-inference/src/lib.rs

@@ -139,6 +139,85 @@ pub fn generate_with_cache<B: Backend>(
     output_ids
 }
 
+/// 流式生成：每生成一个新 token，立即调用 `on_token` 回调输出该 token 的解码文本。
+/// 返回完整的 output token IDs。
+pub fn generate_stream<B: Backend>(
+    model: &LlamaForCausalLM<B>,
+    tokenizer: &TokenizerWrapper,
+    input_ids: &[u32],
+    config: &GenerationConfig,
+    eos_token_id: &EosTokenId,
+    device: &B::Device,
+    mut on_token: impl FnMut(&str),
+) -> Vec<u32> {
+    let mut output_ids = input_ids.to_vec();
+    let mut cache: Option<LlamaKVCache<B>> = None;
+
+    // 第一步：完整 prompt 输入
+    {
+        let input_ints: Vec<i64> = input_ids.iter().map(|&x| x as i64).collect();
+        let input_tensor =
+            Tensor::<B, 1, Int>::from_ints(input_ints.as_slice(), device)
+                .unsqueeze::<2>();
+
+        let (logits, new_cache) = model.forward_with_cache(input_tensor, cache.as_ref());
+        cache = Some(new_cache);
+
+        let next_token = sample_last(&logits, config.temperature, config.top_p);
+        output_ids.push(next_token);
+
+        // 流式输出第一个 token
+        if let Ok(text) = tokenizer.decode(&[next_token], true) {
+            on_token(&text);
+        }
+
+        if eos_token_id.contains(next_token) {
+            return output_ids;
+        }
+    }
+
+    // 后续步骤：每步生成一个 token
+    let mut count = 1;
+    loop {
+        if let Some(max) = config.max_new_tokens {
+            if count >= max {
+                break;
+            }
+        }
+
+        let last_token = output_ids[output_ids.len() - 1];
+        let input_ints: Vec<i64> = vec![last_token as i64];
+        let input_tensor =
+            Tensor::<B, 1, Int>::from_ints(input_ints.as_slice(), device)
+                .unsqueeze::<2>();
+
+        let (logits, new_cache) = model.forward_with_cache(input_tensor, cache.as_ref());
+        cache = Some(new_cache);
+
+        let vocab_size = model.config.vocab_size;
+        let next_token_logits = logits
+            .clone()
+            .slice([0..1, 0..1, 0..vocab_size])
+            .reshape([vocab_size]);
+
+        let next_token = sample(&next_token_logits, config.temperature, config.top_p);
+        output_ids.push(next_token);
+
+        // 流式输出新 token
+        if let Ok(text) = tokenizer.decode(&[next_token], true) {
+            on_token(&text);
+        }
+
+        if eos_token_id.contains(next_token) {
+            break;
+        }
+
+        count += 1;
+    }
+
+    output_ids
+}
+
 fn sample_last<B: Backend>(logits: &Tensor<B, 3, Float>, temperature: f32, top_p: f32) -> u32 {
     let shape = logits.shape();
     let dims: [usize; 3] = shape.dims();
@@ -269,6 +348,29 @@ impl<B: Backend> MiniCPM<B> {
         self.tokenizer.decode(new_ids, true)
     }
 
+    /// 流式生成：每生成一个 token 立即调用 `on_token` 回调，
+    /// 参数为该 token 解码后的文本。
+    pub fn generate_stream(
+        &self,
+        prompt: &str,
+        think: bool,
+        config: &GenerationConfig,
+        on_token: impl FnMut(&str),
+    ) -> anyhow::Result<String> {
+        let input_ids = self.tokenizer.apply_chat_template(prompt, think)?;
+        let output_ids = generate_stream(
+            &self.model,
+            &self.tokenizer,
+            &input_ids,
+            config,
+            &self.config.eos_token_id,
+            &self.device,
+            on_token,
+        );
+        let new_ids = &output_ids[input_ids.len()..];
+        self.tokenizer.decode(new_ids, true)
+    }
+
     pub fn config(&self) -> &LlamaConfig {
         &self.config
     }

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(())