Fix generate(): unpack chat-template BatchEncoding instead of passing positionally

src/text2sql.py

@@ -130,20 +130,26 @@ def generate_sql(
     else:
         messages = [{"role": "user", "content": build_instruction(question, schema, evidence)}]
 
-    inputs = tokenizer.apply_chat_template(
+    # return_dict=True + unpacking is version-robust: apply_chat_template returns a
+    # BatchEncoding (dict) in newer transformers, and passing that positionally to
+    # generate() makes it treat the dict as input_ids and fail on `.shape`.
+    enc = tokenizer.apply_chat_template(
         messages,
         return_tensors="pt",
         add_generation_prompt=True,
-    ).to(model.device)
+        return_dict=True,
+    )
+    enc = {k: v.to(model.device) for k, v in enc.items()}
+    input_len = enc["input_ids"].shape[-1]
 
     with torch.no_grad():
         outputs = model.generate(
-            inputs,
+            **enc,
             max_new_tokens=max_new_tokens if cot else min(max_new_tokens, 256),
             do_sample=False,
             pad_token_id=tokenizer.eos_token_id,
         )
-    raw = tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True)
+    raw = tokenizer.decode(outputs[0][input_len:], skip_special_tokens=True)
 
     if cot:
         return extract_final_sql(raw), raw.strip()