fix: TimesFM 2.5 working end-to-end

.gitignore

@@ -11,3 +11,4 @@ dist/
 scratch/
 *.spec
 data/
+libs/

backend/main.py

@@ -4486,33 +4486,35 @@ def market_status_now() -> List[Dict[str, Any]]:
 
 
 # ──────────────────────────────────────────────────────────────────────────────
-# Kronos Forecaster (identical to v3, with CLIP_DEFAULT fix retained)
+# ──────────────────────────────────────────────────────────────────────────────
+# TimesFM 2.5 Forecaster
 # ──────────────────────────────────────────────────────────────────────────────
 class TimesFMForecaster:
     """Async wrapper around Google TimesFM 2.5 (200M, PyTorch).
 
-    External contract identical to the old KronosForecaster / new TimesFMForecaster:
-        await forecaster.forecast(df, horizon=N, ...) -> {
-            p10, p50, p90,          # np.ndarray of shape (horizon,)
-            model_name, context_length,
-            input_semantics, output_semantics
-        }
+    model.forecast(horizon, inputs) → (point_forecast, quantile_forecast)
+      point_forecast:    np.ndarray (batch, horizon)
+      quantile_forecast: np.ndarray (batch, horizon, 10)
+        index mapping: [mean, q10, q20, q30, q40, q50, q60, q70, q80, q90]
+
     Input:  OHLCV DataFrame; OHLC4 = (O+H+L+C)/4 computed internally.
     Output: p10/p50/p90 future OHLC4 values for `horizon` future bars.
     """
-    MAX_CONTEXT = 15_360          # TimesFM 2.5 supports up to 16 384 ctx
     MODEL_HF_ID = "google/timesfm-2.5-200m-pytorch"
-    # quantile_forecast tensor indices (batch, horizon, 10)
-    # timesfm returns: [mean, q10, q20, q30, q40, q50, q60, q70, q80, q90]
-    _Q10 = 1
-    _Q50 = 5   # median
-    _Q90 = 9
+    # Keep context comfortably within 16384 limit; leaves room for any horizon
+    MAX_CONTEXT   = 8_192
+    MAX_HORIZON   = 300    # hard cap — matches API limit of horizon<=300
+    # quantile axis indices in the (batch, horizon, 10) tensor
+    _Q10 = 1   # 10th percentile
+    _Q50 = 5   # 50th percentile (median)
+    _Q90 = 9   # 90th percentile
 
     def __init__(self) -> None:
         self._model: Optional[Any] = None
         self._loaded = False
         self._lock: Optional[asyncio.Lock] = None
         self._predict_lock: Optional[asyncio.Lock] = None
+        self._compiled_horizon: Optional[int] = None  # last compiled max_horizon
 
     async def _get_lock(self) -> asyncio.Lock:
         if self._lock is None:
@@ -4533,11 +4535,34 @@ class TimesFMForecaster:
         if self._model is None:
             return "not_loaded"
         try:
-            params = list(self._model.parameters())
-            return str(params[0].device) if params else "cpu"
+            return str(self._model.model.device)
         except Exception:
             return "cpu"
 
+    def _compile(self, horizon: int) -> None:
+        """(Re-)compile with a max_horizon that covers `horizon`."""
+        import math
+        output_patch = 128  # TimesFM 2.5 output_patch_len
+        # max_horizon must be a multiple of output_patch_len
+        max_h = math.ceil(horizon / output_patch) * output_patch
+        max_h = max(max_h, output_patch)
+        max_h = min(max_h, self.MAX_HORIZON)
+        # context must be multiple of input_patch_len=32
+        ctx = self.MAX_CONTEXT  # already a multiple of 32
+        self._model.compile(
+            timesfm.ForecastConfig(
+                max_context=ctx,
+                max_horizon=max_h,
+                normalize_inputs=True,
+                use_continuous_quantile_head=True,
+                force_flip_invariance=True,
+                infer_is_positive=True,
+                fix_quantile_crossing=True,
+            )
+        )
+        self._compiled_horizon = max_h
+        logger.info("[TimesFM] Compiled: ctx=%d max_horizon=%d", ctx, max_h)
+
     async def _lazy_load(self) -> None:
         if self._loaded:
             return
@@ -4548,37 +4573,21 @@ class TimesFMForecaster:
             if not TIMESFM_AVAILABLE:
                 raise HTTPException(
                     status_code=503,
-                    detail="TimesFM not installed. Run: pip install timesfm[torch]",
+                    detail="TimesFM not installed. Clone https://github.com/google-research/timesfm and pip install -e .[torch]",
                 )
             try:
-                device = (
-                    "cuda" if torch.cuda.is_available()
-                    else "mps" if (hasattr(torch.backends, "mps") and torch.backends.mps.is_available())
-                    else "cpu"
-                )
-                logger.info("[TimesFM] Loading %s on %s ...", self.MODEL_HF_ID, device)
-
+                logger.info("[TimesFM] Loading %s ...", self.MODEL_HF_ID)
                 model = await asyncio.to_thread(
                     timesfm.TimesFM_2p5_200M_torch.from_pretrained,
                     self.MODEL_HF_ID,
                 )
-                # Apply inference config (must run after loading weights)
-                model.compile(
-                    timesfm.ForecastConfig(
-                        max_context=self.MAX_CONTEXT,
-                        normalize_inputs=True,
-                        use_continuous_quantile_head=True,
-                        force_flip_invariance=True,
-                        infer_is_positive=True,
-                        fix_quantile_crossing=True,
-                    )
-                )
-
+                # Compile once with the default max_horizon
                 self._model = model
+                self._compile(self.MAX_HORIZON)
                 self._loaded = True
                 STARTUP_STATE["timesfm"]["loaded"] = True
-                STARTUP_STATE["timesfm"]["device"] = device
-                logger.info("[TimesFM] Ready on %s", device)
+                STARTUP_STATE["timesfm"]["device"] = self.device
+                logger.info("[TimesFM] Ready on %s", self.device)
             except Exception as ex:
                 STARTUP_STATE["timesfm"]["last_error"] = str(ex)
                 logger.error("[TimesFM] Init failed: %s", ex, exc_info=True)
@@ -4588,13 +4597,14 @@ class TimesFMForecaster:
         self,
         df: pd.DataFrame,
         horizon: int,
-        # legacy kwargs from callers — absorbed and ignored
+        # legacy kwargs — accepted and ignored for API compatibility
         x_timestamp=None,
         y_timestamp=None,
-        sample_count: int = 10,
+        sample_count: int = 0,
     ) -> Dict[str, Any]:
         await self._lazy_load()
         assert self._model is not None
+
         try:
             # Build single OHLC4 series: (O+H+L+C)/4
             ohlc4 = (
@@ -4603,24 +4613,26 @@ class TimesFMForecaster:
                 .astype(np.float32)
                 .values
             )
-
-            # Clip to model context limit (take the most-recent N bars)
             context_len = min(len(ohlc4), self.MAX_CONTEXT)
             ohlc4_ctx = ohlc4[-context_len:]
 
+            # Re-compile if horizon exceeds current compiled max_horizon
+            if self._compiled_horizon is None or horizon > self._compiled_horizon:
+                self._compile(horizon)
+
             t0 = time.time()
             predict_lock = await self._get_predict_lock()
             async with predict_lock:
-                # model.forecast() expects list[np.ndarray]
+                # TimesFM API: forecast(horizon: int, inputs: list[np.ndarray])
                 point_forecast, quantile_forecast = await asyncio.to_thread(
                     self._model.forecast,
-                    inputs=[ohlc4_ctx],
-                    horizon=horizon,
+                    horizon,
+                    [ohlc4_ctx],
                 )
             elapsed = time.time() - t0
             logger.info(
-                "[TimesFM] %.2fs | horizon=%d ctx=%d",
-                elapsed, horizon, context_len,
+                "[TimesFM] %.2fs | horizon=%d ctx=%d device=%s",
+                elapsed, horizon, context_len, self.device,
             )
 
             if torch is not None and torch.cuda.is_available():
@@ -4628,9 +4640,9 @@ class TimesFMForecaster:
 
             # point_forecast:    (1, horizon)
             # quantile_forecast: (1, horizon, 10)
-            p10 = quantile_forecast[0, :, self._Q10].astype(float)
-            p50 = quantile_forecast[0, :, self._Q50].astype(float)
-            p90 = quantile_forecast[0, :, self._Q90].astype(float)
+            p10 = quantile_forecast[0, :horizon, self._Q10].astype(float)
+            p50 = quantile_forecast[0, :horizon, self._Q50].astype(float)
+            p90 = quantile_forecast[0, :horizon, self._Q90].astype(float)
 
             return {
                 "p10": p10,
@@ -4648,7 +4660,7 @@ class TimesFMForecaster:
                     "volume_mode": "omitted",
                     "amount_mode": "omitted",
                     "adapter_mode": "timesfm_native",
-                    "normalization": "timesfm_internal",
+                    "normalization": "timesfm_internal_revin",
                 },
                 "output_semantics": {
                     "forecast_channel": "ohlc4",
@@ -4662,6 +4674,8 @@ class TimesFMForecaster:
             raise HTTPException(status_code=500, detail=f"TimesFM prediction failed: {ex}")
 
 
+
+
 forecaster = TimesFMForecaster()
 
 

run.bat

@@ -1,33 +1,77 @@
 @echo off
-TITLE Kronos AI Trading Terminal v6.0
+TITLE Kronos AI Trading Terminal
 COLOR 0B
+chcp 65001 >nul 2>&1
 
-echo ====================================================
-echo    KRONOS AI TRADING TERMINAL - STARTUP
-echo ====================================================
+echo.
+echo ╔══════════════════════════════════════════════════════╗
+echo ║     KRONOS AI TRADING TERMINAL — STARTUP             ║
+echo ║     Powered by Google TimesFM 2.5 (200M params)      ║
+echo ╚══════════════════════════════════════════════════════╝
+echo.
 
-echo [1/3] Preparing launcher...
-
-:: Check if virtual environment exists
-if not exist "venv\" (
-    echo [ERROR] Virtual environment 'venv' not found.
-    echo Please create it using: python -m venv venv
+:: ── [1/4] Check virtual environment ─────────────────────────────────────────
+echo [1/4] Checking virtual environment...
+if not exist "venv\Scripts\python.exe" (
+    echo [ERROR] Virtual environment not found.
+    echo         Run: python -m venv venv
+    echo         Then: venv\Scripts\pip install -r requirements.txt
     pause
-    exit /b
+    exit /b 1
 )
 
-echo [2/3] Activating Virtual Environment...
+:: ── [2/4] Activate venv ──────────────────────────────────────────────────────
+echo [2/4] Activating virtual environment...
 call venv\Scripts\activate
 
-echo [3/3] Launching AI Trading Terminal...
-echo [INFO] Application will open in your browser automatically.
-echo [INFO] Press CTRL+C in this window to stop the server.
+:: ── [3/4] Ensure TimesFM is installed ───────────────────────────────────────
+echo [3/4] Checking TimesFM dependency...
+venv\Scripts\python.exe -c "import timesfm; timesfm.TimesFM_2p5_200M_torch" >nul 2>&1
+if %ERRORLEVEL% neq 0 (
+    echo [INFO] TimesFM not found. Installing from local clone...
+    if exist "libs\timesfm\pyproject.toml" (
+        echo [INFO] Installing from libs\timesfm ...
+        venv\Scripts\python.exe -m pip install -e "libs\timesfm[torch]" --quiet
+        if %ERRORLEVEL% neq 0 (
+            echo [ERROR] TimesFM install failed from libs\timesfm.
+            pause
+            exit /b 1
+        )
+    ) else (
+        echo [INFO] Cloning TimesFM from GitHub...
+        git clone --depth=1 https://github.com/google-research/timesfm.git libs\timesfm
+        if %ERRORLEVEL% neq 0 (
+            echo [ERROR] git clone failed. Check your internet connection.
+            pause
+            exit /b 1
+        )
+        echo [INFO] Installing TimesFM...
+        venv\Scripts\python.exe -m pip install -e "libs\timesfm[torch]" --quiet
+        if %ERRORLEVEL% neq 0 (
+            echo [ERROR] TimesFM install failed.
+            pause
+            exit /b 1
+        )
+    )
+    echo [OK] TimesFM installed successfully.
+) else (
+    echo [OK] TimesFM is ready.
+)
+
+:: ── [4/4] Launch server ──────────────────────────────────────────────────────
+echo [4/4] Launching Kronos AI Trading Terminal...
+echo.
+echo [INFO] Server will start at http://127.0.0.1:8000
+echo [INFO] Browser will open automatically in ~3 seconds.
+echo [INFO] TimesFM model downloads on FIRST forecast request (~800MB).
+echo [INFO] Press CTRL+C to stop the server.
+echo.
 
-:: Run the app with dynamic port selection
 python -m backend.launcher
 
 if %ERRORLEVEL% neq 0 (
     echo.
-    echo [ERROR] Backend failed to start.
+    echo [ERROR] Server exited with an error.
+    echo         Check the log above for details.
     pause
 )