patch KronosPredictor for torch 2.6 meta tensor safety

Dockerfile

@@ -19,7 +19,7 @@ RUN pip install --user --no-cache-dir --index-url https://download.pytorch.org/w
 # App deps — all pinned to known-compatible versions
 RUN pip install --user --no-cache-dir \
     "gradio[mcp]==5.30.0" \
-    "huggingface_hub>=0.25.0,<0.27.0" \
+    "huggingface_hub>=0.27.0,<1.0" \
     "numpy>=1.26,<2.3" \
     "pandas>=2.1" \
     "yfinance>=0.2.50" \

model/kronos.py

@@ -502,8 +502,24 @@ class KronosPredictor:
         
         self.device = device
 
-        self.tokenizer = self.tokenizer.to(self.device)
-        self.model = self.model.to(self.device)
+        # torch 2.6 + huggingface_hub 0.27+ compat: skip .to() when target is CPU
+        # (models already load to CPU by default) and handle meta tensors gracefully.
+        def _safe_to(mod, dev):
+            try:
+                has_meta = any(p.is_meta for p in mod.parameters()) or \
+                           any(b.is_meta for b in mod.buffers())
+            except Exception:
+                has_meta = False
+            if has_meta:
+                # Should be rare for fresh from_pretrained; skip .to() — torch raises on meta.to()
+                return mod
+            # Skip no-op .to('cpu') to avoid touching the moved-from-meta code path
+            if str(dev) in ("cpu", "torch.device('cpu')") and next(mod.parameters()).device.type == "cpu":
+                return mod
+            return mod.to(dev)
+
+        self.tokenizer = _safe_to(self.tokenizer, self.device)
+        self.model = _safe_to(self.model, self.device)
 
     def generate(self, x, x_stamp, y_stamp, pred_len, T, top_k, top_p, sample_count, verbose):