Fix for Transformers v5

app.py

@@ -66,9 +66,10 @@ def compute_mean_logprob(prompt_text, generated_text):
     # Always use chat template: the model is an instruct model, so
     # logprobs are meaningful only in the chat context.
     message = [{"role": "user", "content": prompt_text}]
-    prompt_ids = tokenizer.apply_chat_template(
+    encoded = tokenizer.apply_chat_template(
         message, add_generation_prompt=True, return_tensors="pt"
-    ).to(model.device)
+    )
+    prompt_ids = (encoded if isinstance(encoded, torch.Tensor) else encoded["input_ids"]).to(model.device)
     gen_ids = tokenizer.encode(
         generated_text, add_special_tokens=False, return_tensors="pt"
     ).to(model.device)

solution.py

@@ -29,7 +29,8 @@ class LaDisparition:
     def __call__(self, prompt, max_tokens=20, beam_width=5):
         # Option 2: we use self.tokenizer.apply_chat_template to tokenize the prompt
         message = [{"role": "user", "content": prompt}]
-        input_ids = self.tokenizer.apply_chat_template(message, add_generation_prompt=True, return_tensors="pt").to(self.model.device)
+        encoded = self.tokenizer.apply_chat_template(message, add_generation_prompt=True, return_tensors="pt")
+        input_ids = (encoded if isinstance(encoded, torch.Tensor) else encoded["input_ids"]).to(self.model.device)
         prompt_len = input_ids.shape[1]
         
         # Beam search: maintain multiple hypotheses
@@ -155,7 +156,8 @@ class ToulouseSequence:
     def __call__(self, prompt, max_tokens=20):
         # Option 2: we use self.tokenizer.apply_chat_template to tokenize the prompt
         message = [{"role": "user", "content": prompt}]
-        inputs = self.tokenizer.apply_chat_template(message, add_generation_prompt=True, return_tensors="pt").to(self.model.device)
+        encoded = self.tokenizer.apply_chat_template(message, add_generation_prompt=True, return_tensors="pt")
+        inputs = (encoded if isinstance(encoded, torch.Tensor) else encoded["input_ids"]).to(self.model.device)
         prompt_length = inputs.shape[1]
         
         # Generate tokens one by one