feat: 更新模型配置和错误处理，添加protobuf依赖

app.py

@@ -1,5 +1,7 @@
 from __future__ import annotations
 
+import logging
+import os
 from dataclasses import dataclass
 from typing import List, Optional
 
@@ -16,6 +18,9 @@ except Exception:  # pragma: no cover
     AutoTokenizer = None
 
 
+logger = logging.getLogger(__name__)
+
+
 @dataclass
 class SummaryOutput:
     summary: str
@@ -24,7 +29,7 @@ class SummaryOutput:
 
 
 class SummarizationConfig:
-    model_name: str = "google/mt5-small"
+    model_name: str = os.getenv("MODEL_NAME", "IDEA-CCNL/Randeng-T5-Char-57M-MultiTask-Chinese")
     max_source_length: int = 1024
     max_target_length: int = 160
     num_beams: int = 4
@@ -85,17 +90,19 @@ class SimpleExtractiveSummarizer:
 
 
 class HybridSummarizer:
-    def __init__(self, model_name: str = "google/mt5-small"):
-        self.model_name = model_name
+    def __init__(self, model_name: str | None = None):
+        self.model_name = os.getenv("MODEL_NAME", model_name or SummarizationConfig.model_name)
         self.backend_name = "fallback"
         self.tokenizer = None
         self.model = None
         self.fallback = SimpleExtractiveSummarizer()
         self.device = "cpu"
+        self.load_error: str | None = None
         self._try_load_transformer()
 
     def _try_load_transformer(self) -> None:
         if AutoTokenizer is None or AutoModelForSeq2SeqLM is None or torch is None:
+            self.load_error = "torch/transformers not installed"
             return
         try:
             self.tokenizer = AutoTokenizer.from_pretrained(self.model_name)
@@ -103,7 +110,10 @@ class HybridSummarizer:
             self.device = "cuda" if torch.cuda.is_available() else "cpu"
             self.model.to(self.device)
             self.backend_name = "transformer"
-        except Exception:
+            self.load_error = None
+        except Exception as exc:
+            self.load_error = f"{type(exc).__name__}: {exc}"
+            logger.exception("Failed to load transformer model: %s", self.model_name)
             self.tokenizer = None
             self.model = None
             self.backend_name = "fallback"
@@ -128,7 +138,7 @@ class HybridSummarizer:
         )
 
     def _summarize_with_transformer(self, text: str, target_length: int | None) -> str:
-        prompt = f"请根据目标长度 {target_length or 120} 字生成摘要：{text}"
+        prompt = f"summarize: {text}"
         inputs = self.tokenizer(
             prompt,
             return_tensors="pt",
@@ -167,7 +177,12 @@ class SummarizeResponse(BaseModel):
 
 @app.get("/health")
 def health():
-    return {"status": "ok", "backend": engine.backend_name}
+    return {
+        "status": "ok",
+        "backend": engine.backend_name,
+        "model_name": engine.model_name,
+        "load_error": engine.load_error,
+    }
 
 
 @app.post("/summarize", response_model=SummarizeResponse)
@@ -277,6 +292,8 @@ def root():
         <div class="card">
           <h1>Transformer Summarizer Demo</h1>
           <p>这是一个基于 Transformer 的中文文本摘要演示系统。你可以通过下面两个按钮进入接口文档或检查服务状态，也可以直接调用摘要接口。</p>
+          <p>当前模型：<code>{engine.model_name}</code></p>
+          <p>当前后端：<code>{engine.backend_name}</code></p>
 
           <div class="btns">
             <a class="btn primary" href="/docs" target="_blank" rel="noreferrer">打开接口文档</a>
@@ -294,6 +311,7 @@ def root():
 }</code></pre>
             <p>4. 点击 <code>Execute</code> 后查看返回的摘要结果。</p>
             <p>5. 如果想确认服务是否正常，可点击 <code>检查服务状态</code>，返回 <code>ok</code> 即表示运行正常。</p>
+            <p>6. 如果健康检查里的 <code>backend</code> 仍然是 <code>fallback</code>，说明 Transformer 模型没有成功加载，请先查看 <code>load_error</code> 的原因。</p>
             <div class="meta">
               提示：如果文本里有换行，请确保是合法 JSON。建议直接在 Swagger 页面提交，避免手写 JSON 出错。
             </div>

requirements.txt

@@ -4,4 +4,4 @@ pydantic>=2.7.0
 transformers>=4.41.0
 sentencepiece>=0.2.0
 torch>=2.1.0
-
+protobuf>=4.25.0