Upload handler.py

handler.py

@@ -124,6 +124,9 @@ class EndpointHandler:
             
             # Generate transcription with anti-hallucination parameters
             with torch.no_grad(), torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.float16):
+                # Add language forcing to inputs instead of generation params
+                model_inputs.update(self.processor.get_decoder_prompt_ids(language="french", task="transcribe"))
+                
                 predicted_ids = self.model.generate(
                     **model_inputs,
                     max_length=max_length,
@@ -136,7 +139,6 @@ class EndpointHandler:
                     length_penalty=1.0,
                     use_cache=True,
                     pad_token_id=self.processor.tokenizer.eos_token_id,
-                    forced_decoder_ids=self.french_decoder_ids,
                     suppress_tokens=[],
                     begin_suppress_tokens=[]
                 )