Spaces:
Running on Zero
Running on Zero
| # -*- coding: utf-8 -*- | |
| """ | |
| Jev-style 判別式決策互動 POC（繁體中文） | |
| 後端：jevos（feder-cr/jev，OpenVINO int8，CPU 可跑） | |
| - 是非題 (noul)：輸入狀態＋問題，回傳 P(yes) | |
| - 多選題 (choice)：輸入狀態＋問題＋候選選項，回傳每選項機率 | |
| - 每次回應標示延遲與 output_tokens=0（零解碼，判別 vs 生成的核心差異） | |
| 免費部署：Hugging Face Spaces（CPU Basic 免費方案即可） | |
| 1. 在 huggingface.co/new-space 建一個 Gradio Space（公開） | |
| 2. 上傳 app.py＋requirements.txt | |
| 3. 首次啟動會自動下載 jevos（約 660MB）並快取於 /data，之後秒啟 | |
| 本地運行：見 README.md | |
| """ | |
| import json | |
| import os | |
| import shutil | |
| import subprocess | |
| import threading | |
| import time | |
| import zipfile | |
| import tarfile | |
| import gradio as gr | |
| import requests | |
| import spaces # ZeroGPU 執行期內建；不可寫進 requirements.txt（會蓋掉內建模組） | |
| BASE = os.path.dirname(os.path.abspath(__file__)) | |
| if os.path.isdir("/data") and os.access("/data", os.W_OK): | |
| DATA = "/data" | |
| _DATA_PERSISTENT = True | |
| else: | |
| DATA = os.path.join(BASE, ".poc-data") | |
| _DATA_PERSISTENT = False | |
| os.makedirs(DATA, exist_ok=True) | |
| BIN_URL = "https://github.com/feder-cr/jev/releases/download/jevos-v3/jev-linux-x64.tar.gz" | |
| MODEL_URL = "https://github.com/feder-cr/jev/releases/download/jevos-v3/jevos-v3-openvino-int8.zip" | |
| 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（支援斷點續傳＋重試），顯示進度百分比。""" | |
| global _last_pct_shown | |
| if os.path.exists(dst): | |
| return dst | |
| name = url.split("/")[-1] | |
| tmp = dst + ".part" | |
| _last_pct_shown = -1 | |
| for attempt in range(1, 4): | |
| try: | |
| start = os.path.getsize(tmp) if os.path.exists(tmp) else 0 | |
| if start > 0: | |
| print(f"[poc] 續傳 {name}（從 {start // 1024 // 1024}MB 繼續）", flush=True) | |
| else: | |
| _set_phase("downloading", f"開始下載 {name}") | |
| headers = {"Range": f"bytes={start}-"} if start else {} | |
| with requests.get(url, headers=headers, stream=True, timeout=30) as r: | |
| r.raise_for_status() | |
| total = int(r.headers.get("content-length", 0) or 0) | |
| if start and r.status_code != 206: | |
| # 伺服器不支援續傳，從頭來 | |
| print("[poc] 伺服器不支援續傳，從頭下載", flush=True) | |
| start, total, mode = 0, total, "wb" | |
| else: | |
| mode = "ab" if start else "wb" | |
| full_total = total + start | |
| downloaded = start | |
| with open(tmp, mode) as f: | |
| for chunk in r.iter_content(chunk_size=1024 * 256): | |
| if not chunk: | |
| continue | |
| f.write(chunk) | |
| downloaded += len(chunk) | |
| if full_total: | |
| pct = min(100, downloaded * 100 // full_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) | |
| os.rename(tmp, dst) | |
| _progress["pct"] = 100 | |
| print("[poc] 下載完成", flush=True) | |
| return dst | |
| except (requests.RequestException, OSError) as e: | |
| print(f"[poc] 下載中斷（嘗試 {attempt}/3）：{type(e).__name__}，5 秒後續傳重試", flush=True) | |
| time.sleep(5) | |
| raise RuntimeError(f"下載失敗，已重試 3 次：{name}") | |
| 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 | |
| with _lock: | |
| if _server is not None and _server.poll() is None: | |
| return | |
| _ensure_backend_locked() | |
| 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 | |
| if backend_ready(): | |
| _set_phase("ready", "jevos 已在運行") | |
| return | |
| try: | |
| jev_dir, binary = _find_or_prepare_backend() | |
| # 保險：確保可執行（git/網頁上傳可能掉執行權限） | |
| try: | |
| os.chmod(binary, 0o755) | |
| except Exception: | |
| pass | |
| _ensure_model_weights(os.path.join(jev_dir, "model")) | |
| _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 _ensure_model_weights(model_dir): | |
| """確保 openvino_model.bin 存在；缺失時從 release 下載 zip 並只解出該檔（支援續傳）。""" | |
| bin_path = os.path.join(model_dir, "openvino_model.bin") | |
| if os.path.isfile(bin_path): | |
| return | |
| print("[poc] 缺少 openvino_model.bin，從 release 下載模型包並解出該檔…", flush=True) | |
| _set_phase("downloading", "下載模型包（約 630MB，只取權重檔，支援續傳）…") | |
| zf = _dl(MODEL_URL, os.path.join(DATA, "jevos-v3-openvino-int8.zip")) | |
| _set_phase("extracting", "解出 openvino_model.bin…") | |
| os.makedirs(model_dir, exist_ok=True) | |
| with zipfile.ZipFile(zf) as z: | |
| for name in z.namelist(): | |
| if name.endswith("openvino_model.bin"): | |
| with z.open(name) as src, open(bin_path, "wb") as dst: | |
| shutil.copyfileobj(src, dst, length=1024 * 1024) | |
| print("[poc] openvino_model.bin 就緒", flush=True) | |
| return | |
| raise RuntimeError("模型包內找不到 openvino_model.bin") | |
| def _find_or_prepare_backend(): | |
| """找後端檔案，順序：repo 內已解壓 → 快取 → repo 內壓縮檔 → 線上下載。 | |
| 回傳 (jev_dir, binary)。jev_dir 下保證有執行檔與 model/（巢狀佈局會建連結）。""" | |
| # 1. 已解壓好的（手動上傳到 repo 根目錄，或之前快取的） | |
| for base in (BASE, DATA): | |
| # 1a. 標準佈局：<base>/jev/{jev(執行檔), model/, *.so} | |
| d = os.path.join(base, "jev") | |
| b = os.path.join(d, "jev") | |
| if os.path.isfile(b) and os.path.isdir(os.path.join(d, "model")): | |
| print(f"[poc] 使用現成後端：{b}", flush=True) | |
| return d, b | |
| # 1b. 巢狀佈局：<base>/jev/jev/{jev(執行檔), *.so}，model/ 在 <base>/jev/model/ | |
| # （tarball 內層目錄被原樣上傳時會長這樣；執行檔在內層，./model 要能被找到） | |
| inner = os.path.join(d, "jev") | |
| b2 = os.path.join(inner, "jev") | |
| outer_model = os.path.join(d, "model") | |
| if os.path.isfile(b2) and os.path.isdir(outer_model): | |
| link = os.path.join(inner, "model") | |
| if not os.path.exists(link): | |
| try: | |
| os.symlink(outer_model, link) | |
| print("[poc] 巢狀佈局：已建立 model 連結", flush=True) | |
| except Exception as e: | |
| print(f"[poc] 建 model 連結失敗：{e}", flush=True) | |
| print(f"[poc] 使用現成後端（巢狀）：{b2}", flush=True) | |
| return inner, b2 | |
| # 2. repo 根目錄有壓縮檔（手動上傳）→ 解到 DATA | |
| jev_dir = os.path.join(DATA, "jev") | |
| tgz = os.path.join(BASE, "jev-linux-x64.tar.gz") | |
| zf = os.path.join(BASE, "jevos-v3-openvino-int8.zip") | |
| if os.path.exists(tgz) and os.path.exists(zf): | |
| print("[poc] 使用 repo 內附壓縮檔解壓…", flush=True) | |
| _set_phase("extracting", "解壓縮 repo 內附檔案…") | |
| os.makedirs(DATA, exist_ok=True) | |
| with tarfile.open(tgz) as tf: | |
| tf.extractall(DATA, filter="data") | |
| print("[poc] 主程式解壓完成，解壓縮模型（約 630MB，需要幾分鐘）…", flush=True) | |
| _extract_zip_progress(zf, jev_dir) | |
| print("[poc] 模型解壓完成", flush=True) | |
| return jev_dir, os.path.join(jev_dir, "jev") | |
| # 3. 線上下載（Spaces 執行期可連外網） | |
| 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, filter="data") | |
| zf = _dl(MODEL_URL, os.path.join(DATA, "jevos-v3-openvino-int8.zip")) | |
| _set_phase("extracting", "解壓縮模型（約 630MB，需要幾分鐘）…") | |
| _extract_zip_progress(zf, jev_dir) | |
| print("[poc] 模型解壓完成", flush=True) | |
| return jev_dir, os.path.join(jev_dir, "jev") | |
| def _extract_zip_progress(zf, jev_dir): | |
| """解壓 zip 並印進度（避免長時間無輸出被誤判卡死）。""" | |
| with zipfile.ZipFile(zf) as z: | |
| members = z.namelist() | |
| total = len(members) | |
| for i, m in enumerate(members): | |
| z.extract(m, jev_dir) | |
| if (i + 1) % 25 == 0 or (i + 1) == total: | |
| print(f"[poc] 模型解壓縮 {i + 1}/{total}", flush=True) | |
| 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", "", "", 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": | |
| payload = {"model": "jev-latest", "state": state, | |
| "questions": {"q1": {"type": "noul", "instructions": question}}} | |
| else: | |
| criteria = {} | |
| for line in options_text.strip().split("\n"): | |
| line = line.strip() | |
| if not line: | |
| continue | |
| if ":" in line: | |
| name, desc = line.split(":", 1) | |
| elif "：" in line: | |
| name, desc = line.split("：", 1) | |
| else: | |
| name, desc = line, "" | |
| name = name.strip() | |
| 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", "", "", status_text() | |
| payload = {"model": "jev-latest", "state": state, | |
| "questions": {"q1": {"type": "choice", "instructions": question, | |
| "criteria": criteria}}} | |
| try: | |
| r = requests.post(API, json=payload, timeout=120) | |
| r.raise_for_status() | |
| data = r.json() | |
| except Exception as 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]}", "", "", status_text() | |
| usage = data.get("usage", {}) | |
| if ans["type"] == "noul": | |
| p = ans["noul"] | |
| bars = f"P(yes) = {p:.3f}\n" + "█" * int(p * 40) + f" {p*100:.1f}%" | |
| summary = f"答案：{'是' if p >= 0.5 else '否'}（信心 {p if p>=0.5 else 1-p:.1%}）" | |
| else: | |
| probs = ans["probabilities"] | |
| bars = "\n".join(f"{k}: {'█' * int(v*40)} {v*100:.1f}%" | |
| for k, v in sorted(probs.items(), key=lambda x: -x[1])) | |
| summary = (f"答案：{ans['choice']}（信心 {ans.get('confidence', 0):.1%}）\n" | |
| 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, 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 = { | |
| "客服訊息分類": ("客人說：「這個月帳單怎麼多了一筆我沒買的東西？」", | |
| "多選題", "這封訊息應該派給哪個團隊？", | |
| "billing: 帳單、付款、退款\nshipping: 寄送、包裹遺失損壞\ntech: 產品故障不會動"), | |
| "是否為退款請求": ("I want my money back for the broken headphones.", | |
| "是非題", "這是一則退款請求嗎？", ""), | |
| "內容審核": ("這家餐廳難吃到我想給一星，老闆還敢嗆客人。", | |
| "是非題", "這則評論含有針對店家的惡意攻擊嗎？", ""), | |
| } | |
| def load_preset(name): | |
| s, m, q, o = PRESETS[name] | |
| return s, m, q, o | |
| def gpu_compat_check(): | |
| """ZeroGPU 硬體要求至少一個 @spaces.GPU 函數才能啟動。 | |
| jevos 為純 CPU 運算（OpenVINO），此函數僅為通過啟動檢查的佔位，實際不參與判別、不消耗有意義的 GPU 配額。""" | |
| return "ZeroGPU 相容檢查通過（判別運算實際於 CPU 執行）" | |
| CSS = """ | |
| .gradio-container { max-width: 900px !important; } | |
| .bar textarea { font-family: monospace !important; white-space: pre !important; } | |
| """ | |
| 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。\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) | |
| gpu_out = gr.Textbox(label="ZeroGPU 相容檢查", value="", 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") | |
| gpu_btn = gr.Button("GPU 相容檢查", 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) | |
| options = gr.Textbox(label="候選選項（多選題用，每行「名稱: 描述」）", lines=4) | |
| btn = gr.Button("送出判別", variant="primary") | |
| out_summary = gr.Textbox(label="判別結果") | |
| out_bars = gr.Textbox(label="機率分佈", elem_classes="bar", lines=6) | |
| 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) | |
| gpu_btn.click(gpu_compat_check, outputs=gpu_out) | |
| btn.click(decide, [state, mode, question, options], | |
| [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__": | |
| print("[poc] app.py v20261003-zerogpu (multi-source backend + @spaces.GPU)", flush=True) | |
| print(f"[poc] 資料目錄：{DATA}（{'重啟後保留，可斷點續傳' if _DATA_PERSISTENT else '重啟後消失，請勿重啟'}）", flush=True) | |
| # Spaces 啟動時先備好後端（背景執行，不擋 UI） | |
| ensure_backend_if_needed() | |
| demo.launch(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)), | |
| css=CSS) | |