fix: unwrap BatchEncoding from apply_chat_template before generate()

app.py

@@ -98,11 +98,23 @@ class GenReq(BaseModel):
 def _prep(req: "GenReq"):
     tok, model, dev = _state["tok"], _state["model"], _state["device"]
     prompt = (req.prompt or "").strip()[:4000]
+    ids = None
     try:  # use the instruct chat template when available
-        ids = tok.apply_chat_template([{"role": "user", "content": prompt}],
-                                      add_generation_prompt=True, return_tensors="pt").to(dev)
+        enc = tok.apply_chat_template([{"role": "user", "content": prompt}],
+                                      add_generation_prompt=True, return_tensors="pt")
+        # Transformers 5.x may hand back a BatchEncoding rather than a bare tensor. Passing that
+        # into generate() dies on `inputs_tensor.shape[0]`, so always unwrap to the tensor.
+        if hasattr(enc, "input_ids"):
+            ids = enc.input_ids
+        elif isinstance(enc, dict):
+            ids = enc["input_ids"]
+        else:
+            ids = enc
     except Exception:
-        ids = tok(prompt, return_tensors="pt").input_ids.to(dev)  # batch_size = 1 ONLY
+        ids = None
+    if ids is None or not hasattr(ids, "shape"):
+        ids = tok(prompt, return_tensors="pt").input_ids
+    ids = ids.to(dev)  # batch_size = 1 ONLY
     temp = float(req.temperature if req.temperature is not None else 0.7)
     gen = dict(
         max_new_tokens=min(int(req.max_new_tokens or MAX_NEW), MAX_NEW),