feat: run Kronos on CPU to avoid GPU conflict with llama-server

predix.py

@@ -329,13 +329,9 @@ def quant(
         threading.Thread(target=start_cli_dash, daemon=True).start()
         time.sleep(1)
 
-    # ---- Kronos Factor: skip if GPU unavailable (CUDA OOM with llama-server) ----
+    # ---- Kronos Factor: CPU inference to avoid GPU conflict with llama-server ----
     try:
-        import torch
-        if torch.cuda.is_available() and torch.cuda.get_device_properties(0).total_memory > 20 * 1024**3:
-            _ensure_kronos_factor_in_pool(console)
-        else:
-            console.print("[dim]Kronos Factor skipped — GPU < 20GB or CUDA unavailable[/dim]")
+        _ensure_kronos_factor_in_pool(console)
     except Exception:
         console.print("[dim]Kronos Factor skipped — torch not available[/dim]")
 

rdagent/components/coder/kronos_adapter.py

@@ -91,7 +91,7 @@ class KronosAdapter:
     TOKENIZER_ID = "NeoQuasar/Kronos-Tokenizer-2k"
 
     def __init__(self, device: Optional[str] = None, max_context: int = 512):
-        self.device = device or ("cuda" if _cuda_available() else "cpu")
+        self.device = device or "cpu"
         self.max_context = max_context
         self._predictor = None
 
@@ -236,7 +236,7 @@ def build_kronos_factor(
     Returns:
         MultiIndex (datetime, instrument) DataFrame with column "KronosPredReturn".
     """
-    device = device or ("cuda" if _cuda_available() else "cpu")
+    device = device or "cpu"
     logger.info(f"Loading data from {hdf5_path}...")
     raw = pd.read_hdf(hdf5_path, key="data")
 
@@ -312,7 +312,7 @@ def evaluate_kronos_model(
     Returns:
         dict with keys: IC_mean, IC_std, IC_IR (IC / std), hit_rate, n_predictions
     """
-    device = device or ("cuda" if _cuda_available() else "cpu")
+    device = device or "cpu"
     raw = pd.read_hdf(hdf5_path, key="data")
     instrument = raw.index.get_level_values("instrument").unique()[0]
     df = raw.xs(instrument, level="instrument")