Update handler.py

handler.py

@@ -42,7 +42,6 @@ class EndpointHandler:
 
             # compute decoder_input_ids for french
             forced_ids = self.processor.get_decoder_prompt_ids(language="french", task="transcribe")
-            # convert to tensor [ [id1,id2,...] ]
             self.french_decoder_input_ids = torch.tensor(
                 [[tok_id for _, tok_id in forced_ids]],
                 device="cuda" if torch.cuda.is_available() else "cpu"
@@ -98,8 +97,7 @@ class EndpointHandler:
             with torch.no_grad(), torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.float16):
                 predicted_ids = self.model.generate(
                     **model_inputs,
-                    decoder_input_ids=self.french_decoder_input_ids,  # ✅ remplace forced_decoder_ids
-                    forced_decoder_ids=None,                          # ✅ évite le conflit
+                    decoder_input_ids=self.french_decoder_input_ids,  # ✅ seul forçage
                     max_length=max_length,
                     num_beams=num_beams,
                     temperature=temperature,