fix: torch 2.6, chronos inputs=, timesfm fallback, yfinance curl_cffi+retry

Dockerfile

@@ -14,7 +14,7 @@ WORKDIR /home/user/app
 
 # CPU-only torch from PyTorch index (skip CUDA wheels, ~200MB vs ~2GB)
 RUN pip install --user --no-cache-dir --index-url https://download.pytorch.org/whl/cpu \
-    torch==2.4.1
+    torch==2.6.0
 
 # App deps — all pinned to known-compatible versions
 RUN pip install --user --no-cache-dir \
@@ -23,6 +23,7 @@ RUN pip install --user --no-cache-dir \
     "numpy>=1.26,<2.3" \
     "pandas>=2.1" \
     "yfinance>=0.2.50" \
+    "curl_cffi>=0.7" \
     "websockets>=13.0" \
     "einops>=0.7" \
     "safetensors>=0.4" \

app.py

@@ -22,20 +22,50 @@ import gradio as gr
 # -----------------------------------------------------------------------------
 
 def _load_ohlc(symbol: str, lookback_days: int) -> pd.DataFrame:
-    df = yf.download(symbol, period=f"{lookback_days + 10}d",
-                     interval="1d", progress=False, auto_adjust=False)
-    if df is None or df.empty:
-        return pd.DataFrame()
-    if isinstance(df.columns, pd.MultiIndex):
-        df.columns = df.columns.get_level_values(0)
-    df = df.reset_index()
-    df.columns = [str(c).lower() for c in df.columns]
-    if "date" not in df.columns:
+    def _tidy(df):
+        if df is None or df.empty:
+            return pd.DataFrame()
+        if isinstance(df.columns, pd.MultiIndex):
+            df.columns = df.columns.get_level_values(0)
+        df = df.reset_index()
+        df.columns = [str(c).lower() for c in df.columns]
+        if "date" not in df.columns and "datetime" in df.columns:
+            df["date"] = df["datetime"]
+        if "date" not in df.columns:
+            return pd.DataFrame()
+        df["timestamps"] = pd.to_datetime(df["date"])
+        keep = ["timestamps", "open", "high", "low", "close", "volume"]
+        df = df[[c for c in keep if c in df.columns]].dropna().tail(lookback_days).reset_index(drop=True)
+        return df
+
+    # Primary: yf.download. Fallback: yf.Ticker().history(). Uses curl_cffi chrome impersonation if available.
+    session = None
+    try:
+        from curl_cffi import requests as cureq
+        session = cureq.Session(impersonate="chrome")
+    except Exception:
+        session = None
+
+    period = f"{lookback_days + 10}d"
+    try:
+        df = yf.download(symbol, period=period, interval="1d",
+                         progress=False, auto_adjust=False,
+                         session=session) if session else \
+             yf.download(symbol, period=period, interval="1d",
+                         progress=False, auto_adjust=False)
+        out = _tidy(df)
+        if len(out) >= 32:
+            return out
+    except Exception:
+        pass
+
+    # Fallback path
+    try:
+        t = yf.Ticker(symbol, session=session) if session else yf.Ticker(symbol)
+        df = t.history(period=period, interval="1d", auto_adjust=False)
+        return _tidy(df)
+    except Exception:
         return pd.DataFrame()
-    df["timestamps"] = pd.to_datetime(df["date"])
-    keep = ["timestamps", "open", "high", "low", "close", "volume"]
-    df = df[[c for c in keep if c in df.columns]].dropna().tail(lookback_days).reset_index(drop=True)
-    return df
 
 
 def _direction(pct: float) -> int:
@@ -130,7 +160,7 @@ def forecast_chronos(symbol: str, lookback_days: int = 180, pred_days: int = 30)
             return {"status": "error", "error": f"insufficient data for {symbol}", "n_lookback": len(kdf)}
         context = torch.tensor(kdf["close"].values, dtype=torch.float32)
         quantiles, mean = _get_chronos().predict_quantiles(
-            context=context, prediction_length=pred_days, quantile_levels=[0.1, 0.5, 0.9],
+            inputs=context, prediction_length=pred_days, quantile_levels=[0.1, 0.5, 0.9],
         )
         median = quantiles[0, :, 1].cpu().numpy()
         low = quantiles[0, :, 0].cpu().numpy()
@@ -163,7 +193,16 @@ def _get_timesfm():
     global _timesfm
     if _timesfm is None:
         import timesfm
-        _timesfm = timesfm.TimesFm_2p5_200M_torch.from_pretrained(TIMESFM_MODEL_ID)
+        # TimesFM package API varies across versions. Try the 2.5-specific class first,
+        # then fall back to the generic TimesFm(hparams, checkpoint) constructor.
+        cls = getattr(timesfm, "TimesFm_2p5_200M_torch", None)
+        if cls is not None:
+            _timesfm = cls.from_pretrained(TIMESFM_MODEL_ID)
+        else:
+            _timesfm = timesfm.TimesFm(
+                hparams=timesfm.TimesFmHparams(backend="torch"),
+                checkpoint=timesfm.TimesFmCheckpoint(huggingface_repo_id=TIMESFM_MODEL_ID),
+            )
     return _timesfm