Update app.py

app.py

@@ -1,3 +1,38 @@
+# -------------------------------
+# 4. 本地推論模型設定
+# -------------------------------
+MODEL_MAP = {
+    "Auto": None,  # 自動選擇
+    "Gemma-2B": "google/gemma-2b",
+    "Gemma-7B": "google/gemma-7b",
+    "BTLM-3B-8K": "cerebras/btlm-3b-8k",
+    "gpt-oss-20B": "openai-community/gpt-oss-20b"
+}
+
+# 快取 pipeline 避免每次重建
+_loaded_pipelines = {}
+
+def get_pipeline(model_name):
+    if model_name not in _loaded_pipelines:
+        print(f"🔄 正在載入模型 {model_name} ...")
+        model_id = MODEL_MAP[model_name]
+        generator = pipeline(
+            "text-generation",
+            model=model_id,
+            tokenizer=model_id,
+            device_map="auto",
+        )
+        _loaded_pipelines[model_name] = generator
+    return _loaded_pipelines[model_name]
+
+def call_local_inference(model_name, prompt, max_new_tokens=512):
+    try:
+        generator = get_pipeline(model_name)
+        outputs = generator(prompt, max_new_tokens=max_new_tokens, do_sample=True, temperature=0.7)
+        return outputs[0]["generated_text"]
+    except Exception as e:
+        return f"（生成失敗：{e}）"
+
 # -------------------------------
 # 5. 生成文章（即時進度）
 # -------------------------------
@@ -66,3 +101,4 @@ with gr.Blocks() as demo:
 # -------------------------------
 if __name__ == "__main__":
     demo.launch()
+