Update app.py

app.py

@@ -15,7 +15,6 @@ Jev-style 判別式決策互動 POC（繁體中文）
 import json
 import os
 import subprocess
-import sys
 import threading
 import time
 import urllib.request
@@ -35,17 +34,71 @@ API = "http://127.0.0.1:8017/v1/systemone"
 _server = None
 _lock = threading.Lock()
 
+# 後端狀態（給 UI 顯示用，避免首次下載時使用者以為當機）
+_progress = {"phase": "idle", "pct": 0, "msg": ""}
+_last_pct_shown = -1
+
+
+def _set_phase(phase, msg="", pct=None):
+    global _last_pct_shown
+    _progress["phase"] = phase
+    _progress["msg"] = msg
+    if pct is not None:
+        _progress["pct"] = pct
+    print(f"[poc] {msg or phase}", flush=True)
+
+
+def status_text():
+    ph = _progress["phase"]
+    if ph == "ready":
+        return "✅ 後端就緒（jevos 運行中）"
+    if ph == "downloading":
+        return f"⏳ 下載模型中 {_progress['pct']}%（首次啟動需下載約 660MB，請稍候）"
+    if ph == "extracting":
+        return "⏳ 解壓縮模型中…"
+    if ph == "starting":
+        return "⏳ 啟動 jevos 中（載入模型約需 10 秒）…"
+    if ph == "error":
+        return f"❌ 後端錯誤：{_progress['msg']}"
+    return "⏳ 後端準備中…"
+
 
 def _dl(url, dst):
+    """下載檔案，顯示進度百分比。"""
+    global _last_pct_shown
     if os.path.exists(dst):
         return dst
-    print(f"[poc] 下載 {url}", flush=True)
+    _set_phase("downloading", f"開始下載 {url.split('/')[-1]}")
     tmp = dst + ".part"
-    urllib.request.urlretrieve(url, tmp)
+    _last_pct_shown = -1
+
+    def hook(blocks, block_size, total):
+        global _last_pct_shown
+        if total and total > 0:
+            pct = min(100, blocks * block_size * 100 // total)
+            _progress["pct"] = pct
+            if pct // 10 != _last_pct_shown // 10:
+                _last_pct_shown = pct
+                print(f"[poc] 下載進度 {pct}%", flush=True)
+
+    urllib.request.urlretrieve(url, tmp, reporthook=hook)
     os.rename(tmp, dst)
+    _progress["pct"] = 100
+    print("[poc] 下載完成", flush=True)
     return dst
 
 
+def backend_ready():
+    """不阻塞地檢查 jevos 是否已在服務。"""
+    try:
+        requests.post(API, json={"model": "jev-latest", "state": "ping",
+                                 "questions": {"q": {"type": "noul", "instructions": "ok?"}}},
+                      timeout=3)
+        return True
+    except Exception:
+        return False
+
+
 def ensure_backend():
     """下載、解壓、啟動 jevos（冪等，可重複呼叫）。"""
     global _server
@@ -58,45 +111,45 @@ def ensure_backend():
 def _ensure_backend_locked():
     global _server
     # 先看是不是已經有 server 在跑（例如本機手動啟動的）→ 有就直接用，不下載
-    try:
-        requests.post(API, json={"model": "jev-latest", "state": "ping",
-                                 "questions": {"q": {"type": "noul", "instructions": "ok?"}}},
-                      timeout=5)
-        print("[poc] jevos 已在運行", flush=True)
+    if backend_ready():
+        _set_phase("ready", "jevos 已在運行")
         return
-    except Exception:
-        pass
-    jev_dir = os.path.join(DATA, "jev")
-    binary = os.path.join(jev_dir, "jev")
-    model_dir = os.path.join(jev_dir, "model")
-    if not (os.path.exists(binary) and os.path.isdir(model_dir)):
-        os.makedirs(DATA, exist_ok=True)
-        tgz = _dl(BIN_URL, os.path.join(DATA, "jev-linux-x64.tar.gz"))
-        with tarfile.open(tgz) as tf:
-            tf.extractall(DATA)
-        zf = _dl(MODEL_URL, os.path.join(DATA, "jevos-v3-openvino-int8.zip"))
-        with zipfile.ZipFile(zf) as z:
-            z.extractall(jev_dir)  # 內含 model/
-        os.chmod(binary, 0o755)
-    print("[poc] 啟動 jevos…", flush=True)
-    _server = subprocess.Popen([binary, "serve"], cwd=jev_dir,
-                               stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
-    for _ in range(60):
-        time.sleep(1)
-        try:
-            requests.post(API, json={"model": "jev-latest", "state": "ping",
-                                     "questions": {"q": {"type": "noul", "instructions": "ok?"}}},
-                          timeout=5)
-            print("[poc] jevos 就緒", flush=True)
-            return
-        except Exception:
-            continue
-    raise RuntimeError("jevos 啟動逾時")
+    try:
+        jev_dir = os.path.join(DATA, "jev")
+        binary = os.path.join(jev_dir, "jev")
+        model_dir = os.path.join(jev_dir, "model")
+        if not (os.path.exists(binary) and os.path.isdir(model_dir)):
+            os.makedirs(DATA, exist_ok=True)
+            tgz = _dl(BIN_URL, os.path.join(DATA, "jev-linux-x64.tar.gz"))
+            _set_phase("extracting", "解壓縮 jevos 主程式…")
+            with tarfile.open(tgz) as tf:
+                tf.extractall(DATA)
+            zf = _dl(MODEL_URL, os.path.join(DATA, "jevos-v3-openvino-int8.zip"))
+            _set_phase("extracting", "解壓縮模型（約 630MB）…")
+            with zipfile.ZipFile(zf) as z:
+                z.extractall(jev_dir)  # 內含 model/
+            os.chmod(binary, 0o755)
+        _set_phase("starting", "啟動 jevos…")
+        _server = subprocess.Popen([binary, "serve"], cwd=jev_dir,
+                                   stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+        for _ in range(90):
+            time.sleep(1)
+            if backend_ready():
+                _set_phase("ready", "jevos 就緒")
+                return
+        raise RuntimeError("jevos 啟動逾時（90 秒）")
+    except Exception as e:
+        _set_phase("error", f"{type(e).__name__}: {e}")
+        raise
 
 
 def decide(state, mode, question, options_text):
     if not (state or "").strip() or not (question or "").strip():
-        return "請先輸入狀態文字與問題", "", ""
+        return "請先輸入狀態文字與問題", "", "", status_text()
+    # 後端還沒好就不要卡住：直接告訴使用者目前進度
+    if not backend_ready():
+        ensure_backend_if_needed()
+        return status_text() + "，請稍後再按一次「送出判別」。", "", "", status_text()
     ensure_backend()
     t0 = time.perf_counter()
     if mode == "是非題":
@@ -118,7 +171,7 @@ def decide(state, mode, question, options_text):
             desc = desc.strip() or name  # 無描述時用名稱代替（choice 需要 criteria 描述）
             criteria[name] = desc
         if not criteria:
-            return "請至少輸入一個候選選項", "", ""
+            return "請至少輸入一個候選選項", "", "", status_text()
         payload = {"model": "jev-latest", "state": state,
                    "questions": {"q1": {"type": "choice", "instructions": question,
                                        "criteria": criteria}}}
@@ -127,12 +180,12 @@ def decide(state, mode, question, options_text):
         r.raise_for_status()
         data = r.json()
     except Exception as e:
-        return f"後端呼叫失敗：{type(e).__name__}（{e}）", "", ""
+        return f"後端呼叫失敗：{type(e).__name__}（{e}）", "", "", status_text()
     ms = (time.perf_counter() - t0) * 1000
     try:
         ans = data["answers"]["q1"]
     except (KeyError, TypeError):
-        return f"後端回傳格式異常：{str(data)[:300]}", "", ""
+        return f"後端回傳格式異常：{str(data)[:300]}", "", "", status_text()
     usage = data.get("usage", {})
     if ans["type"] == "noul":
         p = ans["noul"]
@@ -146,7 +199,13 @@ def decide(state, mode, question, options_text):
                    f"各選項機率：{json.dumps(probs, ensure_ascii=False)}")
     meta = (f"延遲 {ms:.0f} ms ｜ input_tokens={usage.get('input_tokens','?')} ｜ "
             f"output_tokens={usage.get('output_tokens','?')}（零解碼）")
-    return summary, bars, meta
+    return summary, bars, meta, status_text()
+
+
+def ensure_backend_if_needed():
+    """確保背景下載執行緒有在跑（不阻塞）。"""
+    t = threading.Thread(target=ensure_backend, daemon=True)
+    t.start()
 
 
 PRESETS = {
@@ -173,9 +232,12 @@ CSS = """
 with gr.Blocks(title="Jev-style 判別式決策 POC") as demo:
     gr.Markdown("# ⚡ Jev-style 判別式決策 POC\n"
                 "輸入一段**狀態文字**＋**問題**，模型直接輸出**機率分佈**——不生成任何文字（`output_tokens=0`）。"
-                "這就是 System 1 判別引擎：傳統 LLM Agent 每步 1–3 秒還可能寫壞 JSON，這裡一次前向傳播就給出完整機率分佈。")
+                "這就是 System 1 判別引擎：傳統 LLM Agent 每步 1–3 秒還可能寫壞 JSON，這裡一次前向傳播就給出完整機率分佈。\n\n"
+                "> ⏰ **首次啟動**需要下載模型（約 660MB），請看下方「後端狀態」等它就緒再按送出，之後啟動都是秒開。")
+    status_box = gr.Textbox(label="後端狀態", value="⏳ 後端準備中…", interactive=False)
     with gr.Row():
         preset = gr.Dropdown(list(PRESETS.keys()), label="範例一鍵填入", value="客服訊息分類")
+        refresh_btn = gr.Button("重新整理狀態", size="sm")
     state = gr.Textbox(label="狀態文字（state）", lines=3)
     mode = gr.Radio(["是非題", "多選題"], label="題型", value="多選題")
     question = gr.Textbox(label="問題（instructions）", lines=1)
@@ -186,12 +248,14 @@ with gr.Blocks(title="Jev-style 判別式決策 POC") as demo:
     out_meta = gr.Textbox(label="效能", lines=1)
 
     preset.change(load_preset, preset, [state, mode, question, options])
+    refresh_btn.click(lambda: status_text(), outputs=status_box)
     btn.click(decide, [state, mode, question, options],
-              [out_summary, out_bars, out_meta])
+              [out_summary, out_bars, out_meta, status_box])
     demo.load(load_preset, preset, [state, mode, question, options])
+    demo.load(lambda: status_text(), outputs=status_box)
 
 if __name__ == "__main__":
     # Spaces 啟動時先備好後端（背景執行，不擋 UI）
-    threading.Thread(target=ensure_backend, daemon=True).start()
+    ensure_backend_if_needed()
     demo.launch(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)),
                 css=CSS)