Sync from GitHub refs/heads/main @ fa80e51

Dockerfile

@@ -38,11 +38,14 @@ ENV PORT=7860
 ENV HF_HOME=/tmp/hf_cache
 ENV TORCH_HOME=/tmp/torch_cache
 ENV TRANSFORMERS_OFFLINE=0
-# Production warm-start — preload chat (gemma-4-e4b) at lifespan boot so the
-# first /chat request doesn't pay ~30s cold-load. Default in main.py is lazy
-# (opt-in) because Apple Silicon dev hangs when Gemma + FashionSigLIP
-# cohabit MPS unified memory; HF Spaces L40S has no such constraint.
+# Production warm-start — preload chat (gemma-4-e4b) + trend_predict (TimesFM 2.5)
+# at lifespan boot so first /chat or /api/v1/ml/trends/predict request doesn't pay
+# ~30s / ~10s cold-load. Defaults in main.py are lazy (opt-in) because Apple Silicon
+# dev hangs when Gemma + FashionSigLIP cohabit MPS unified memory; HF Spaces L40S
+# has no such constraint. TimesFM 2.5 also requires .compile(ForecastConfig) which
+# runs at first load — preload absorbs that one-time JIT cost.
 ENV ML_PRELOAD_CHAT=1
+ENV ML_PRELOAD_TIMESFM=1
 
 EXPOSE 7860
 USER appuser

config/models.yaml

@@ -33,6 +33,13 @@ models:
     provider: huggingface
     model_id: "amazon/chronos-bolt-tiny"
 
+  # TimesFM 2.5：Google 200M 參數，Apache 2.0，univariate forecasting
+  # 取代 chronos-bolt-tiny（9M）做主力時序預測。chronos-bolt 保留 entry
+  # 給 TREND_MODEL=chronos-bolt 回退（git-bisect / smoke prod issue）。
+  timesfm-2.5:
+    provider: huggingface
+    model_id: "google/timesfm-2.5-200m-pytorch"
+
   xgboost-valuation:
     provider: xgboost
     model_path: "models/valuation_xgb.joblib"
@@ -181,7 +188,8 @@ capabilities:
     min_samples: 50
 
   trend_predict:
-    model: chronos-bolt
+    # ${TREND_MODEL:-timesfm-2.5}：預設 TimesFM 2.5；設 TREND_MODEL=chronos-bolt 可回退舊 9M tiny。
+    model: "${TREND_MODEL:-timesfm-2.5}"
     dev_mode: "${TREND_DEV_MODE:-false}"
 
   size_recommend:

pyproject.toml

@@ -19,6 +19,7 @@ dependencies = [
     "scikit-learn>=1.4.0",
     "joblib>=1.3.0",
     "chronos-forecasting>=2.0.0",
+    "timesfm>=1.3.0",
     "pytrends>=4.9.0",
     "timm>=1.0.25",
     "einops>=0.8.2",

src/main.py

@@ -74,7 +74,7 @@ async def lifespan(_app: FastAPI) -> AsyncIterator[None]:
     except Exception:
         pass  # Non-fatal: will lazy-load on first request
 
-    # 3. 背景預載 chat capability (gemma-4-e4b)，不阻塞 server 啟動
+    # 3. 背景預載 capability，不阻塞 server 啟動
     #    載入完成前 health endpoint 仍可回應，model 相關 endpoint 會等 lock 釋放
     #
     #    預設 lazy（不預載）— Gemma 4 ~7GB 與 FashionSigLIP 共住在 Apple
@@ -83,15 +83,23 @@ async def lifespan(_app: FastAPI) -> AsyncIterator[None]:
     #    /chat 等請求才拉權重；只 SigLIP 工作流（backfill / paste / embed）
     #    永遠不會觸發 cohabitation。
     #
-    #    ML_PRELOAD_CHAT=1 才預載 — 給 HF Spaces L40S production 用，避免
-    #    第一個聊天請求等 ~30s 冷啟。
+    #    Production 預載（HF Spaces L40S）：
+    #    - ML_PRELOAD_CHAT=1     → 預載 chat (gemma-4-e4b)，避免第一個聊天請求 ~30s 冷啟
+    #    - ML_PRELOAD_TIMESFM=1  → 預載 trend_predict (TimesFM 2.5)，避免第一個趨勢請求
+    #                              ~10s 冷啟 + ForecastConfig.compile() 開銷
+    _preload_caps: list[str] = []
     if os.environ.get("ML_PRELOAD_CHAT") == "1":
+        _preload_caps.append("chat")
+    if os.environ.get("ML_PRELOAD_TIMESFM") == "1":
+        _preload_caps.append("trend_predict")
+
+    if _preload_caps:
 
         async def _preload_critical_models() -> None:
             from src.registry import get_registry
 
             reg = get_registry()
-            for cap in ["chat"]:
+            for cap in _preload_caps:
                 try:
                     await reg.get(cap)
                     # WARN level so HF Spaces / Render container logs surface
@@ -123,7 +131,7 @@ async def lifespan(_app: FastAPI) -> AsyncIterator[None]:
     else:
         # WARN level — see Pre-loaded note above re: root logger default.
         _logger.warning(
-            "ml-service startup: ML_PRELOAD_CHAT not set — chat capability lazy-loads on first request"
+            "ml-service startup: no ML_PRELOAD_* set — all capabilities lazy-load on first request"
         )
 
     yield  # 應用程式運行中

src/providers/huggingface.py

@@ -75,10 +75,14 @@ class HuggingFaceProvider(ModelProvider):
         if "florence" in model_id.lower():
             return await self._load_florence(parsed)
 
-        # Chronos-Bolt（時序預測）
+        # Chronos-Bolt（時序預測，舊；保留給 TREND_MODEL=chronos-bolt 回退）
         if "chronos" in model_id.lower():
             return await self._load_chronos(parsed)
 
+        # TimesFM 2.5（Google 200M 時序基礎模型，Apache 2.0）
+        if "timesfm" in model_id.lower():
+            return await self._load_timesfm(parsed)
+
         # 預設：Causal LM（Qwen 系列）+ optional LoRA
         return await self._load_causal_lm(parsed)
 
@@ -359,6 +363,53 @@ class HuggingFaceProvider(ModelProvider):
             config={"model_id": config.model_id, "device": device},
         )
 
+    async def _load_timesfm(self, config: HFModelConfig) -> LoadedModel:
+        """載入 TimesFM 2.5 — 在 worker thread 中執行避免阻塞 event loop。"""
+        return await asyncio.to_thread(self._load_timesfm_sync, config)
+
+    def _load_timesfm_sync(self, config: HFModelConfig) -> LoadedModel:
+        """同步載入 Google TimesFM 2.5 200M（由 worker thread 呼叫）。
+
+        【為什麼需要 .compile()？】
+        TimesFM 2.5 的 from_pretrained 只 load weights，inference 前必須先 compile
+        ForecastConfig — 否則 forecast() 會 RuntimeError。compile 一次後可重用。
+
+        【ForecastConfig 參數選擇】
+        - max_horizon=128：90 天足夠，留 buffer 給未來 180 天 endpoint 不需重 compile
+        - max_context=1024：歷史價格序列上限，遠超 price_history 實際需求
+        - normalize_inputs：消除品牌間絕對價差影響
+        - use_continuous_quantile_head + fix_quantile_crossing：取得單調遞增 quantile
+        - infer_is_positive：價格非負強制 forecast≥0
+        - force_flip_invariance：抗 sign-flip pretraining artifact
+        """
+        import torch
+        import timesfm
+
+        device = self._resolve_device(config.device)
+        logger.info("Loading TimesFM 2.5 (%s) on %s", config.model_id, device)
+
+        torch.set_float32_matmul_precision("high")
+        model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(config.model_id)
+        model.compile(
+            timesfm.ForecastConfig(
+                max_context=1024,
+                max_horizon=128,
+                normalize_inputs=True,
+                use_continuous_quantile_head=True,
+                force_flip_invariance=True,
+                infer_is_positive=True,
+                fix_quantile_crossing=True,
+            )
+        )
+        self._loaded = True
+        logger.info("TimesFM 2.5 loaded + compiled successfully")
+
+        return LoadedModel(
+            model=model,
+            tokenizer=None,
+            config={"model_id": config.model_id, "device": device},
+        )
+
     @staticmethod
     def _resolve_device(device_str: str) -> str:
         """解析裝置字串，auto 會偵測 cuda > mps > cpu 順序。"""

src/services/trend_predictor.py

@@ -38,10 +38,30 @@ logger = logging.getLogger(__name__)
 # ============================================================================
 # 模式切換
 # ============================================================================
-# 預設 false = 生產模式（用真實 Chronos + Google Trends）
+# 預設 false = 生產模式（用真實 TimesFM + Google Trends）
 # 設 true = 開發模式（回傳假資料，不打外部 API）
 DEV_MODE = os.environ.get("TREND_DEV_MODE", "false").lower() == "true"
 
+# 主力模型切換：預設 TimesFM 2.5（200M, Apache 2.0）
+# 設 TREND_MODEL=chronos-bolt 可回退舊 chronos-bolt-tiny（9M）做 git-bisect
+TREND_MODEL = os.environ.get("TREND_MODEL", "timesfm-2.5").lower()
+
+# TimesFM 3-gate 信心門檻
+# n<32: 純規則 fallback（mock + low_confidence）
+# 32 ≤ n < 100: TimesFM + low_confidence flag
+# n ≥ 100: TimesFM 全信心
+TIMESFM_LOW_CONFIDENCE_MIN = 32
+TIMESFM_FULL_CONFIDENCE_MIN = 100
+
+_DISCLAIMER_INSUFFICIENT = (
+    "Insufficient price history (<32 records) — forecast uses brand baseline "
+    "rules, not the TimesFM model."
+)
+_DISCLAIMER_LIMITED = (
+    "Limited price history (32-100 records) — TimesFM forecast may have wider "
+    "uncertainty."
+)
+
 # ============================================================================
 # ModelRegistry 整合
 # ============================================================================
@@ -99,6 +119,10 @@ class TrendPrediction(BaseModel):
     predicted_price_90d: float  # 90 天後的預測均價
     google_trend_score: float  # Google Trends 熱度分數（0-100）
     forecast_generated_at: str  # 預測產生時間（ISO 格式）
+    # ── 新增：信心揭露（PR-2 TimesFM 2.5 swap，additive 不破壞既有 caller）──
+    # low_confidence=true 時前端應顯示 disclaimer（n<100 或 fallback）
+    low_confidence: bool = False
+    disclaimer: str = ""
 
 
 # ============================================================================
@@ -182,14 +206,22 @@ def predict_trend(
     參數：
       brand: 品牌名（如 "Gucci"）
       price_history: 歷史成交紀錄，格式 [{"date": "2024-01-15", "price": 1200.0}, ...]
-                     可選——沒有的話，Chronos 會 fallback 到 mock 價格 + 真實 Google Trends
+                     可選——沒有的話，model 會 fallback 到 mock 價格 + 真實 Google Trends
 
     回傳：
-      TrendPrediction — 含 30d/90d 預測價、趨勢方向、Google Trends 分數
+      TrendPrediction — 含 30d/90d 預測價、趨勢方向、Google Trends 分數、
+                        low_confidence flag、disclaimer
+
+    分派邏輯：
+      - DEV_MODE=true → mock（不打外部 API）
+      - TREND_MODEL=chronos-bolt → 舊 Chronos-Bolt-Tiny（git-bisect rollback path）
+      - 預設 TREND_MODEL=timesfm-2.5 → Google TimesFM 2.5 200M（3-gate 信心邏輯）
     """
     if DEV_MODE:
         return _mock_predict(brand)
-    return _chronos_predict(brand, price_history)
+    if TREND_MODEL == "chronos-bolt":
+        return _chronos_predict(brand, price_history)
+    return _timesfm_predict(brand, price_history)
 
 
 # ============================================================================
@@ -197,15 +229,22 @@ def predict_trend(
 # ============================================================================
 
 
-def _mock_predict(brand: str) -> TrendPrediction:
+def _mock_predict(
+    brand: str,
+    low_confidence: bool = False,
+    disclaimer: str = "",
+) -> TrendPrediction:
     """
-    開發模式：用硬編碼的基準價 + 趨勢方向，模擬預測結果。
+    開發模式 / fallback：用硬編碼的基準價 + 趨勢方向，模擬預測結果。
 
     不打任何外部 API，適合離線開發和 CI 測試。
     價格變化邏輯：
     - UP:     30d +3%, 90d +8%
     - DOWN:   30d -3%, 90d -8%
     - STABLE: 30d +0.5%, 90d +1%
+
+    當生產 path 因資料不足/模型失敗 fallback 到此函式時，caller 應傳
+    low_confidence=True + disclaimer 字串，讓前端能正確標示。
     """
     baseline = _BRAND_BASELINES.get(brand, 500.0)
     direction = _MOCK_DIRECTIONS.get(brand, TrendDirection.STABLE)
@@ -228,6 +267,8 @@ def _mock_predict(brand: str) -> TrendPrediction:
         predicted_price_90d=price_90d,
         google_trend_score=trend_score,
         forecast_generated_at=datetime.now(tz=timezone.utc).isoformat(),
+        low_confidence=low_confidence,
+        disclaimer=disclaimer,
     )
 
 
@@ -393,4 +434,152 @@ def _chronos_predict(
         # 任何錯誤（模型載入失敗、tensor 格式錯誤等）都 fallback 到 mock
         # 確保 API 不會因為預測失敗而整個掛掉
         logger.exception("Chronos prediction failed for brand=%s", brand)
-        return _mock_predict(brand)
+        return _mock_predict(
+            brand,
+            low_confidence=True,
+            disclaimer=_DISCLAIMER_INSUFFICIENT,
+        )
+
+
+# ============================================================================
+# 生���模式：TimesFM 2.5（主力）
+# ============================================================================
+# Google Research 200M 參數 decoder-only 時序基礎模型，Apache 2.0。
+# 取代 chronos-bolt-tiny（9M）作為主力預測。
+
+TIMESFM_MODEL_ID = "google/timesfm-2.5-200m-pytorch"
+
+# Lazy loading：第一次呼叫載入 + compile，之後重複使用
+_timesfm_model = None
+
+
+def _load_timesfm():
+    """
+    載入並 compile TimesFM 2.5。第一次呼叫後快取於全域變數。
+
+    【為什麼要 .compile()？】
+    TimesFM 2.5 from_pretrained 只 load weights；inference 前必須先 compile
+    ForecastConfig — 否則 forecast() 會 RuntimeError。compile 一次後可重用。
+
+    【ForecastConfig 參數】
+    max_horizon=128：90d 足夠，buffer 留 180d endpoint 不需重 compile
+    max_context=1024：歷史價格序列上限，遠超實際 use case
+    normalize_inputs：消除品牌間絕對價差影響
+    use_continuous_quantile_head + fix_quantile_crossing：取得單調遞增 quantile
+    infer_is_positive：價格非負強制 forecast≥0
+    force_flip_invariance：抗 sign-flip pretraining artifact
+    """
+    global _timesfm_model
+    if _timesfm_model is not None:
+        return _timesfm_model
+
+    import timesfm
+
+    logger.info("Loading TimesFM 2.5 model (%s)...", TIMESFM_MODEL_ID)
+    torch.set_float32_matmul_precision("high")
+    model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(TIMESFM_MODEL_ID)
+    model.compile(
+        timesfm.ForecastConfig(
+            max_context=1024,
+            max_horizon=128,
+            normalize_inputs=True,
+            use_continuous_quantile_head=True,
+            force_flip_invariance=True,
+            infer_is_positive=True,
+            fix_quantile_crossing=True,
+        )
+    )
+    _timesfm_model = model
+    logger.info("TimesFM 2.5 loaded + compiled on %s", DEVICE)
+    return _timesfm_model
+
+
+def _timesfm_predict(
+    brand: str, price_history: list[dict[str, float | str]] | None = None
+) -> TrendPrediction:
+    """
+    用 TimesFM 2.5 做 zero-shot 時序預測，3-gate 信心邏輯。
+
+    【3-gate】
+    n<32：純規則 fallback（mock baseline + low_confidence + disclaimer）
+    32 ≤ n < 100：TimesFM + low_confidence=True
+    n ≥ 100：TimesFM 全信心（low_confidence=False）
+
+    【相比 Chronos-Bolt-Tiny 的改進】
+    - 200M params vs 9M：對長序列關係建模能力大幅提升
+    - 連續 quantile head：信賴區間連續而非離散
+    - flip-invariance + is_positive 約束：價格序列特殊化處理
+    - max_context=1024：可吃更長歷史不需 truncate
+    """
+    try:
+        # Step 1: 即時抓 Google Trends 分數（與 Chronos path 行為一致）
+        google_score = _fetch_google_trend_score(brand)
+        n = len(price_history) if price_history else 0
+
+        # Step 2: n<32 純規則 fallback
+        if n < TIMESFM_LOW_CONFIDENCE_MIN:
+            mock = _mock_predict(
+                brand,
+                low_confidence=True,
+                disclaimer=_DISCLAIMER_INSUFFICIENT,
+            )
+            mock.google_trend_score = google_score
+            return mock
+
+        # Step 3: 準備時間序列（按日期排序，取價格 1-D array）
+        # n<32 gate 保證 price_history 非 None；此 raise 是 unreachable 防禦，
+        # 但比 `assert price_history is not None` 安全：python -O 會 strip
+        # asserts，只剩此 explicit raise 才能 narrow 型別 + 保護 None.sorted()
+        # 跑進 except 變成 silent fallback。
+        if price_history is None:
+            raise ValueError("price_history must be non-None past n<32 gate")
+        sorted_history = sorted(price_history, key=lambda x: str(x["date"]))
+        prices = [float(p["price"]) for p in sorted_history]
+
+        # Step 4: TimesFM forecast
+        # inputs 接受 list of 1-D arrays，回傳 (point_forecast=median, quantile_forecast)
+        # point_forecast shape: (batch=1, horizon=90)
+        # 用 numpy float32 直接傳入（避免 torch tensor 型別不符）
+        import numpy as np
+
+        prices_array = np.array(prices, dtype=np.float32)
+        model = _load_timesfm()
+        point_forecast, _quantile_forecast = model.forecast(
+            horizon=90,
+            inputs=[prices_array],
+        )
+        pred_30d = float(point_forecast[0, 29])  # 第 30 天
+        pred_90d = float(point_forecast[0, 89])  # 第 90 天
+
+        # Step 5: 趨勢方向（與 Chronos path 邏輯一致）
+        current_price = prices[-1]
+        if pred_90d > current_price * 1.05:
+            direction = TrendDirection.UP
+        elif pred_90d < current_price * 0.95:
+            direction = TrendDirection.DOWN
+        else:
+            direction = TrendDirection.STABLE
+
+        # Step 6: 信心 flag
+        low_confidence = n < TIMESFM_FULL_CONFIDENCE_MIN
+        disclaimer = _DISCLAIMER_LIMITED if low_confidence else ""
+
+        return TrendPrediction(
+            brand=brand,
+            trend_direction=direction,
+            predicted_price_30d=round(pred_30d, 2),
+            predicted_price_90d=round(pred_90d, 2),
+            google_trend_score=google_score,
+            forecast_generated_at=datetime.now(tz=timezone.utc).isoformat(),
+            low_confidence=low_confidence,
+            disclaimer=disclaimer,
+        )
+    except Exception:
+        # 任何錯誤（模型載入失敗、numpy 格式錯誤等）fallback 到 mock + low_confidence
+        # 確保 API 不會因為預測失敗而整個掛掉
+        logger.exception("TimesFM prediction failed for brand=%s", brand)
+        return _mock_predict(
+            brand,
+            low_confidence=True,
+            disclaimer=_DISCLAIMER_INSUFFICIENT,
+        )

uv.lock

@@ -19,6 +19,15 @@ resolution-markers = [
 [manifest]
 overrides = [{ name = "transformers", specifier = ">=5.5.0" }]
 
+[[package]]
+name = "absl-py"
+version = "2.4.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/64/c7/8de93764ad66968d19329a7e0c147a2bb3c7054c554d4a119111b8f9440f/absl_py-2.4.0.tar.gz", hash = "sha256:8c6af82722b35cf71e0f4d1d47dcaebfff286e27110a99fc359349b247dfb5d4", size = 116543, upload-time = "2026-01-28T10:17:05.322Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/18/a6/907a406bb7d359e6a63f99c313846d9eec4f7e6f7437809e03aa00fa3074/absl_py-2.4.0-py3-none-any.whl", hash = "sha256:88476fd881ca8aab94ffa78b7b6c632a782ab3ba1cd19c9bd423abc4fb4cd28d", size = 135750, upload-time = "2026-01-28T10:17:04.19Z" },
+]
+
 [[package]]
 name = "accelerate"
 version = "1.13.0"
@@ -549,6 +558,19 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/2a/09/f8d8f8f31e4483c10a906437b4ce31bdf3d6d417b73fe33f1a8b59e34228/einops-0.8.2-py3-none-any.whl", hash = "sha256:54058201ac7087911181bfec4af6091bb59380360f069276601256a76af08193", size = 65638, upload-time = "2026-01-26T04:13:18.546Z" },
 ]
 
+[[package]]
+name = "einshape"
+version = "1.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "absl-py" },
+    { name = "numpy" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/cb/c6/95ad0a036656aec1cb32177a5d5abfcfbf53a01c1416484cacb8c7332a84/einshape-1.0.tar.gz", hash = "sha256:53538d75dd099f4ead4a4f786fafdcb0b729bb587e0b3afeca25ceef18c9ac14", size = 14571, upload-time = "2022-12-19T17:09:34.618Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/71/bb/34cc02f13b438d550e4709216ee1df9da8e55e15b0cc87a2cb5dee19a729/einshape-1.0-py3-none-any.whl", hash = "sha256:42da4c2dea3a27f87ee45a7cee5072a636b97cb184bb07bf5d6412ba0ff7b965", size = 21392, upload-time = "2022-12-19T17:09:32.904Z" },
+]
+
 [[package]]
 name = "fastapi"
 version = "0.135.1"
@@ -705,6 +727,30 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/ab/6e/81d47999aebc1b155f81eca4477a616a70f238a2549848c38983f3c22a82/ftfy-6.3.1-py3-none-any.whl", hash = "sha256:7c70eb532015cd2f9adb53f101fb6c7945988d023a085d127d1573dc49dd0083", size = 44821, upload-time = "2024-10-26T00:50:33.425Z" },
 ]
 
+[[package]]
+name = "gitdb"
+version = "4.0.12"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "smmap" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/72/94/63b0fc47eb32792c7ba1fe1b694daec9a63620db1e313033d18140c2320a/gitdb-4.0.12.tar.gz", hash = "sha256:5ef71f855d191a3326fcfbc0d5da835f26b13fbcba60c32c21091c349ffdb571", size = 394684, upload-time = "2025-01-02T07:20:46.413Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a0/61/5c78b91c3143ed5c14207f463aecfc8f9dbb5092fb2869baf37c273b2705/gitdb-4.0.12-py3-none-any.whl", hash = "sha256:67073e15955400952c6565cc3e707c554a4eea2e428946f7a4c162fab9bd9bcf", size = 62794, upload-time = "2025-01-02T07:20:43.624Z" },
+]
+
+[[package]]
+name = "gitpython"
+version = "3.1.47"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "gitdb" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/c1/bd/50db468e9b1310529a19fce651b3b0e753b5c07954d486cba31bbee9a5d5/gitpython-3.1.47.tar.gz", hash = "sha256:dba27f922bd2b42cb54c87a8ab3cb6beb6bf07f3d564e21ac848913a05a8a3cd", size = 216978, upload-time = "2026-04-22T02:44:44.059Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f2/c5/a1bc0996af85757903cf2bf444a7824e68e0035ce63fb41d6f76f9def68b/gitpython-3.1.47-py3-none-any.whl", hash = "sha256:489f590edfd6d20571b2c0e72c6a6ac6915ee8b8cd04572330e3842207a78905", size = 209547, upload-time = "2026-04-22T02:44:41.271Z" },
+]
+
 [[package]]
 name = "google-api-core"
 version = "2.30.0"
@@ -1327,6 +1373,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/7e/82/69e539c4c2027f1e1697e09aaa2449243085a0edf81ae2c6341e84d769b6/multiprocess-0.70.19-py39-none-any.whl", hash = "sha256:0d4b4397ed669d371c81dcd1ef33fd384a44d6c3de1bd0ca7ac06d837720d3c5", size = 133477, upload-time = "2026-01-19T06:47:38.619Z" },
 ]
 
+[[package]]
+name = "narwhals"
+version = "2.20.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/e9/f3/257adc69a71011b4c8cda321b00f02c5bf1980ae38ffd05a58d9632d4de8/narwhals-2.20.0.tar.gz", hash = "sha256:c10994975fa7dc5a68c2cffcddbd5908fc8ebb2d463c5bab085309c0ee1f551e", size = 627848, upload-time = "2026-04-20T12:11:45.427Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/d0/69/f24d3d1c38ad69e256138b4ec2452a8c7cf66be49dc214771ae99dd4f0a0/narwhals-2.20.0-py3-none-any.whl", hash = "sha256:16e750ea5507d4ba6e8d03455b5f93a535e0405976561baea235bca5dc9f475d", size = 449373, upload-time = "2026-04-20T12:11:43.596Z" },
+]
+
 [[package]]
 name = "networkx"
 version = "3.6.1"
@@ -1746,6 +1801,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/f2/26/c56ce33ca856e358d27fda9676c055395abddb82c35ac0f593877ed4562e/pillow-12.1.1-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:cb9bb857b2d057c6dfc72ac5f3b44836924ba15721882ef103cecb40d002d80e", size = 7029880, upload-time = "2026-02-11T04:23:04.783Z" },
 ]
 
+[[package]]
+name = "platformdirs"
+version = "4.9.6"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/9f/4a/0883b8e3802965322523f0b200ecf33d31f10991d0401162f4b23c698b42/platformdirs-4.9.6.tar.gz", hash = "sha256:3bfa75b0ad0db84096ae777218481852c0ebc6c727b3168c1b9e0118e458cf0a", size = 29400, upload-time = "2026-04-09T00:04:10.812Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/75/a6/a0a304dc33b49145b21f4808d763822111e67d1c3a32b524a1baf947b6e1/platformdirs-4.9.6-py3-none-any.whl", hash = "sha256:e61adb1d5e5cb3441b4b7710bea7e4c12250ca49439228cc1021c00dcfac0917", size = 21348, upload-time = "2026-04-09T00:04:09.463Z" },
+]
+
 [[package]]
 name = "pluggy"
 version = "1.6.0"
@@ -2646,6 +2710,19 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/07/39/338d9219c4e87f3e708f18857ecd24d22a0c3094752393319553096b98af/scipy-1.17.1-cp314-cp314t-win_arm64.whl", hash = "sha256:200e1050faffacc162be6a486a984a0497866ec54149a01270adc8a59b7c7d21", size = 25489165, upload-time = "2026-02-23T00:22:29.563Z" },
 ]
 
+[[package]]
+name = "sentry-sdk"
+version = "2.58.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "certifi" },
+    { name = "urllib3" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/26/b3/fb8291170d0e844173164709fc0fa0c221ed75a5da740c8746f2a83b4eb1/sentry_sdk-2.58.0.tar.gz", hash = "sha256:c1144d947352d54e5b7daa63596d9f848adf684989c06c4f5a659f0c85a18f6f", size = 438764, upload-time = "2026-04-13T17:23:26.265Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/fa/eb/d875669993b762556ae8b2efd86219943b4c0864d22204d622a9aee3052b/sentry_sdk-2.58.0-py2.py3-none-any.whl", hash = "sha256:688d1c704ddecf382ea3326f21a67453d4caa95592d722b7c780a36a9d23109e", size = 460919, upload-time = "2026-04-13T17:23:24.675Z" },
+]
+
 [[package]]
 name = "setuptools"
 version = "82.0.1"
@@ -2673,6 +2750,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/b7/ce/149a00dd41f10bc29e5921b496af8b574d8413afcd5e30dfa0ed46c2cc5e/six-1.17.0-py2.py3-none-any.whl", hash = "sha256:4721f391ed90541fddacab5acf947aa0d3dc7d27b2e1e8eda2be8970586c3274", size = 11050, upload-time = "2024-12-04T17:35:26.475Z" },
 ]
 
+[[package]]
+name = "smmap"
+version = "5.0.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/1f/ea/49c993d6dfdd7338c9b1000a0f36817ed7ec84577ae2e52f890d1a4ff909/smmap-5.0.3.tar.gz", hash = "sha256:4d9debb8b99007ae47165abc08670bd74cb74b5227dda7f643eccc4e9eb5642c", size = 22506, upload-time = "2026-03-09T03:43:26.1Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c1/d4/59e74daffcb57a07668852eeeb6035af9f32cbfd7a1d2511f17d2fe6a738/smmap-5.0.3-py3-none-any.whl", hash = "sha256:c106e05d5a61449cf6ba9a1e650227ecfb141590d2a98412103ff35d89fc7b2f", size = 24390, upload-time = "2026-03-09T03:43:24.361Z" },
+]
+
 [[package]]
 name = "starlette"
 version = "0.52.1"
@@ -2707,6 +2793,26 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/32/d5/f9a850d79b0851d1d4ef6456097579a9005b31fea68726a4ae5f2d82ddd9/threadpoolctl-3.6.0-py3-none-any.whl", hash = "sha256:43a0b8fd5a2928500110039e43a5eed8480b918967083ea48dc3ab9f13c4a7fb", size = 18638, upload-time = "2025-03-13T13:49:21.846Z" },
 ]
 
+[[package]]
+name = "timesfm"
+version = "1.3.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "absl-py" },
+    { name = "einshape" },
+    { name = "huggingface-hub" },
+    { name = "numpy" },
+    { name = "pandas" },
+    { name = "scikit-learn" },
+    { name = "typer" },
+    { name = "utilsforecast" },
+    { name = "wandb" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/1d/38/b98e0e45408348ef91ba8b3035db247200d3ad8a8c8eaf62a82f2add7cf8/timesfm-1.3.0.tar.gz", hash = "sha256:c494775c17d373d8d7f2e5562468fdff371c123d1d6e43dcfe92fbf9a86fec4e", size = 48227, upload-time = "2025-07-06T22:07:07.839Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/d9/04/afb1888ab8db240576811b240cc7b7687c348516a72bfd618c5bd98ea9d2/timesfm-1.3.0-py3-none-any.whl", hash = "sha256:a99e22b13654bdbfaecb14a48e4a8dc6c990b191a0322a5cacd09eb0df3bfebe", size = 55019, upload-time = "2025-07-06T22:07:06.366Z" },
+]
+
 [[package]]
 name = "timm"
 version = "1.0.25"
@@ -2973,6 +3079,21 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/39/08/aaaad47bc4e9dc8c725e68f9d04865dbcb2052843ff09c97b08904852d84/urllib3-2.6.3-py3-none-any.whl", hash = "sha256:bf272323e553dfb2e87d9bfd225ca7b0f467b919d7bbd355436d3fd37cb0acd4", size = 131584, upload-time = "2026-01-07T16:24:42.685Z" },
 ]
 
+[[package]]
+name = "utilsforecast"
+version = "0.2.15"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "narwhals" },
+    { name = "numpy" },
+    { name = "packaging" },
+    { name = "pandas" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/d9/55/8a37bb9ce36541fd353466259a07ccfdfaf25c996f3a71d989af4d4c7ba4/utilsforecast-0.2.15.tar.gz", hash = "sha256:c36d65d698a88d0fadc93d2d6737c304c3776397c60ae551ee17aa678caf3659", size = 60609, upload-time = "2025-12-03T16:29:08.652Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/6f/11/6c6ee61958b8e60f634b39e2f9a004f5d1c479cb962a2001fc3c72ceed78/utilsforecast-0.2.15-py3-none-any.whl", hash = "sha256:4b43bf5107e3cba13604cd86e93b5cf4906b57105b1900ccf98b8978aabd4150", size = 40344, upload-time = "2025-12-03T16:29:07.144Z" },
+]
+
 [[package]]
 name = "uvicorn"
 version = "0.41.0"
@@ -3035,6 +3156,35 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/e4/16/c1fd27e9549f3c4baf1dc9c20c456cd2f822dbf8de9f463824b0c0357e06/uvloop-0.22.1-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:6cde23eeda1a25c75b2e07d39970f3374105d5eafbaab2a4482be82f272d5a5e", size = 4296730, upload-time = "2025-10-16T22:17:00.744Z" },
 ]
 
+[[package]]
+name = "wandb"
+version = "0.26.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "click" },
+    { name = "gitpython" },
+    { name = "packaging" },
+    { name = "platformdirs" },
+    { name = "protobuf" },
+    { name = "pydantic" },
+    { name = "pyyaml" },
+    { name = "requests" },
+    { name = "sentry-sdk" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/6a/a4/72a6640e1f566e81f184a426e3e45298d4c6672664de41adb7eb6f64370a/wandb-0.26.1.tar.gz", hash = "sha256:eef2dbaea06f0b1c0cdc5d76f544ae4c2b8848fc512442a00bd59f0502fc8aa1", size = 42159814, upload-time = "2026-04-23T16:27:34.033Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/8c/09/3296235f3906e904f06f2df29eed4d672fb23c0932c9486e2af64f2f2a66/wandb-0.26.1-py3-none-macosx_12_0_arm64.whl", hash = "sha256:2955fe190c005fb83ee6d73f066c8a33f09f3212a1f2eb53faa6581440e456be", size = 24857204, upload-time = "2026-04-23T16:26:58.576Z" },
+    { url = "https://files.pythonhosted.org/packages/a1/ad/e39ca3086534129e42208ba00ed2c6247ce425f890219eeec33b4f162864/wandb-0.26.1-py3-none-macosx_12_0_x86_64.whl", hash = "sha256:55d91cabde98162d7116a5e19ddd052bd9848556243f1da4cbb9ffb7ad435bfc", size = 26014649, upload-time = "2026-04-23T16:27:02.559Z" },
+    { url = "https://files.pythonhosted.org/packages/56/af/400d84a3bdce0b062b4baa70acb6becd2c8018697f4fbf5af9a9e1e406e5/wandb-0.26.1-py3-none-manylinux_2_28_aarch64.whl", hash = "sha256:7c78bc2454cfe1ffa1c3a256060a387356eed8a4488e024d9d2eba8f2b5bd51d", size = 25421317, upload-time = "2026-04-23T16:27:06.411Z" },
+    { url = "https://files.pythonhosted.org/packages/7b/e9/b4bf8f3509dcea1cec52233a38991459654635b5a8e6a494eb912e1b9cfb/wandb-0.26.1-py3-none-manylinux_2_28_x86_64.whl", hash = "sha256:a2c8eeec8706dcd2872e69c3b4d20ec523082fdb4440295491556e219ad2aa67", size = 27192831, upload-time = "2026-04-23T16:27:10.308Z" },
+    { url = "https://files.pythonhosted.org/packages/62/cf/4a6dce0c782223ef0eeea7139daee73418a7322befcf083512c31cebaa18/wandb-0.26.1-py3-none-musllinux_1_2_aarch64.whl", hash = "sha256:2fa768ee0636a569afb7541cf996e56309c47070566a38916823f94e02afe586", size = 25593326, upload-time = "2026-04-23T16:27:14.259Z" },
+    { url = "https://files.pythonhosted.org/packages/df/99/58c3d8c36ae8e2b7d70bf6493eb5daa1cca0231a04b025717b4cd1a78f1e/wandb-0.26.1-py3-none-musllinux_1_2_x86_64.whl", hash = "sha256:5854928725cfeff1f284d5c043cd353f810e5da02eead2c120ef5056ad026fea", size = 27535542, upload-time = "2026-04-23T16:27:18.473Z" },
+    { url = "https://files.pythonhosted.org/packages/7c/d0/4e846ffc1d0cc435518dfa581ce73ac82cfd0ebbf35f3853c9277f632e5f/wandb-0.26.1-py3-none-win32.whl", hash = "sha256:5c2bd44e575ae9944e2764d1aaa031461178276bf2636d5558399c2816ef5cfe", size = 24968151, upload-time = "2026-04-23T16:27:22.086Z" },
+    { url = "https://files.pythonhosted.org/packages/e3/9b/487413eaccefdb58799a226726e24b486e9192d2671c75a4550c160aba23/wandb-0.26.1-py3-none-win_amd64.whl", hash = "sha256:5817785467d3f1676f1812ec19a89f77f6e56dfe67d9f47080075af95f705d3e", size = 24968155, upload-time = "2026-04-23T16:27:25.731Z" },
+    { url = "https://files.pythonhosted.org/packages/04/dc/5baf3e99b3eeb709d6f75124b5bec8cb73d4b38d2b10df7fdcfde4966200/wandb-0.26.1-py3-none-win_arm64.whl", hash = "sha256:f848b7744f896bc04cabbb28360a2814d1551a91fa2c456243e06435729c8a2e", size = 22912416, upload-time = "2026-04-23T16:27:29.456Z" },
+]
+
 [[package]]
 name = "wardrobe-os-ml-service"
 version = "0.0.1"
@@ -3062,6 +3212,7 @@ dependencies = [
     { name = "pyyaml" },
     { name = "replicate" },
     { name = "scikit-learn" },
+    { name = "timesfm" },
     { name = "timm" },
     { name = "torch" },
     { name = "torchvision" },
@@ -3103,6 +3254,7 @@ requires-dist = [
     { name = "pyyaml", specifier = ">=6.0.3" },
     { name = "replicate", specifier = ">=1.0.0" },
     { name = "scikit-learn", specifier = ">=1.4.0" },
+    { name = "timesfm", specifier = ">=1.3.0" },
     { name = "timm", specifier = ">=1.0.25" },
     { name = "torch", specifier = ">=2.5.0" },
     { name = "torchvision", specifier = ">=0.20.0" },