Patch app.py: fix messages format + YAML + queue

app.py

@@ -2,11 +2,11 @@
 import os, torch, gradio as gr
 from transformers import AutoTokenizer, AutoModelForCausalLM
 
-TITLE     = os.getenv("SPACE_TITLE", "LanguageBridge — Multimodal Chatbot (Mistral-7B)")
-MODEL_ID  = os.getenv("MODEL_ID", "aciang/mistral7b-tk-sft-20251019-merged")
+TITLE    = os.getenv("SPACE_TITLE", "LanguageBridge — Multimodal Chatbot (Mistral-7B)")
+MODEL_ID = os.getenv("MODEL_ID", "aciang/mistral7b-tk-sft-20251019-merged")
+
 SYSTEM_PROMPT = os.getenv("SYSTEM_PROMPT",
-    "你是語言橋助教，回答需：條理清晰、步驟簡潔、可連結部落知識與科學方法；必要時列點。"
-)
+    "你是語言橋助教，回答務必：條列、準確、可引用部落知識與科學方法；必要時提供清楚步驟。")
 
 USE_IMAGE = os.getenv("USE_IMAGE", "0") in ("1","true","True")
 USE_AUDIO = os.getenv("USE_AUDIO", "0") in ("1","true","True")
@@ -14,99 +14,70 @@ USE_AUDIO = os.getenv("USE_AUDIO", "0") in ("1","true","True")
 def load_llm():
     dtype = torch.float16 if torch.cuda.is_available() else torch.float32
     try:
-        model = AutoModelForCausalLM.from_pretrained(
-            MODEL_ID, torch_dtype=dtype, device_map="auto"
-        )
+        model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype=dtype, device_map="auto")
     except Exception as e:
-        print(f"[Fallback CPU] {e}")
-        model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype=torch.float32)
+        print(f"[LLM fallback->CPU fp32] {e}")
+        model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype=torch.float32, device_map=None)
     tok = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=True)
-    if tok.pad_token is None:
-        tok.pad_token = tok.eos_token
+    if tok.pad_token is None: tok.pad_token = tok.eos_token
     tok.padding_side = "left"
     return tok, model
 
-tokenizer, llm = load_llm(); llm.eval()
-
-def norm_messages(msgs):
-    """保證為 [{'role': 'user'|'assistant', 'content': '文字'}]"""
-    out = []
-    if not msgs: 
-        return out
-    for m in msgs:
-        if isinstance(m, dict) and 'role' in m and 'content' in m:
-            # Gradio 可能把 content 做成 {'text':...} 或 list；這裡統一成 str
-            c = m['content']
-            if isinstance(c, dict) and 'text' in c:
-                c = c['text']
-            elif isinstance(c, list) and len(c)>0:
-                first = c[0]
-                c = first.get('text', str(first))
-            out.append({'role': m['role'], 'content': str(c)})
-        elif isinstance(m, tuple) and len(m)==2:
-            out.append({'role': m[0], 'content': str(m[1])})
-        else:
-            # 無法辨識就丟棄
-            continue
-    return out
-
-def build_prompt(messages):
-    """把歷史訊息組成簡潔 prompt"""
-    msgs = norm_messages(messages)
-    parts = [SYSTEM_PROMPT, ""]
-    for m in msgs:
-        who = "使用者" if m['role']=="user" else "助教"
-        parts.append(f"{who}：{m['content']}")
-    parts.append("助教：")
-    return "\n".join(parts)
+tokenizer, llm = load_llm()
+llm.eval()
 
 @torch.inference_mode()
-def respond(history, user_text, mx, temp, top_p):
-    # history 為 messages 格式
-    history = norm_messages(history)
-    if user_text and user_text.strip():
-        history.append({'role':'user', 'content': user_text.strip()})
-    prompt = build_prompt(history)
+def generate_reply(history, image, audio, max_new_tokens, temperature, top_p):
+    # history 需為 list[ (role:str, {"content": str}) ]
+    user_text = ""
+    if history:
+        for role, msg in history[::-1]:
+            if role == "user":
+                user_text = msg.get("content","") if isinstance(msg, dict) else str(msg)
+                break
+    prompt = f"{SYSTEM_PROMPT}\n\n使用者：{user_text}\n助教："
     inputs = tokenizer(prompt, return_tensors="pt").to(llm.device)
     out = llm.generate(
         **inputs,
-        max_new_tokens=int(mx),
-        temperature=float(temp),
+        max_new_tokens=int(max_new_tokens),
+        temperature=float(temperature),
         top_p=float(top_p),
-        do_sample=True if float(temp) > 0 else False,
+        do_sample=True,
         eos_token_id=tokenizer.eos_token_id,
-        pad_token_id=tokenizer.pad_token_id
+        pad_token_id=tokenizer.pad_token_id,
     )
     text = tokenizer.decode(out[0], skip_special_tokens=True)
-    if "助教：" in text:
-        text = text.split("助教：",1)[-1].strip()
-    # 回填 assistant 訊息
-    history.append({'role':'assistant', 'content': text})
-    # gr.Chatbot(type="messages") 需要 messages（而不是 HTML）
-    # 清空輸入框
-    return history, ""
+    if "助教：" in text: text = text.split("助教：",1)[-1].strip()
+    return text
 
 with gr.Blocks(title=TITLE) as demo:
-    gr.Markdown(f"## {TITLE}\n- 模型：`{MODEL_ID}`\n- 介面已修正為 Gradio v4 messages 格式")
+    gr.Markdown(f"## {TITLE}\n模型：`{MODEL_ID}`（messages 格式已修正）")
 
-    chat = gr.Chatbot(label="Chatbot", type="messages", height=480)
     with gr.Row():
-        user_txt = gr.Textbox(label="你的問題 / 指令", placeholder="請輸入文字…", scale=4)
-        send = gr.Button("送出 🚀", variant="primary", scale=1)
-        clear = gr.Button("清除", scale=1)
+        chat = gr.Chatbot(height=460, type="messages", show_copy_button=True)
+        with gr.Column(scale=0):
+            user_txt = gr.Textbox(label="你的問題 / 指令", placeholder="請輸入文字…", interactive=True)
+            mx  = gr.Slider(64, 1024, value=256, step=32, label="max_new_tokens")
+            tp  = gr.Slider(0.0, 1.0, value=0.2, step=0.05, label="temperature")
+            top = gr.Slider(0.5, 1.0, value=0.95, step=0.01, label="top_p")
+            go  = gr.Button("送出 🚀", variant="primary")
+            clr = gr.Button("清除")
 
-    with gr.Row():
-        mx   = gr.Slider(64, 1024, value=512, step=32, label="max_new_tokens")
-        temp = gr.Slider(0.0, 1.0, value=0.4,  step=0.05, label="temperature")
-        top  = gr.Slider(0.5, 1.0, value=0.95, step=0.01, label="top_p")
+    def respond(history, text, mx, tp, top):
+        history = history or []
+        history.append(("user", {"content": text}))
+        try:
+            ans = generate_reply(history, None, None, mx, tp, top)
+        except Exception as e:
+            ans = f"(推理失敗：{e})"
+        history.append(("assistant", {"content": ans}))
+        return history, ""
 
-    send.click(respond, inputs=[chat, user_txt, mx, temp, top], outputs=[chat, user_txt])
-    clear.click(lambda: ([], ""), outputs=[chat, user_txt])
+    go.click(respond, inputs=[chat, user_txt, mx, tp, top], outputs=[chat, user_txt])
+    clr.click(lambda: ([], ""), outputs=[chat, user_txt])
 
-    # 不要傳入不支援的 queue 參數（concurrency_count 等）
-    demo.queue()
-    # 在 Space 執行時由服務端啟動，這裡不呼叫 launch()
+    # 重要：不要使用舊參數 concurrency_count（會 Runtime error）
+    demo.queue(max_size=32, api_open=False)
 
 if __name__ == "__main__":
-    # 在本機測試才會啟動；Space 端由 HF 自行啟動
     demo.launch(share=False, server_name="0.0.0.0", server_port=7860, show_error=True)