Remove debug info

Signed-off-by: ziyixiong-nv <[email protected]>

Author: ziyixiong-nv

tensorrt_llm/_torch/pyexecutor/model_engine.py

@@ -1857,7 +1857,6 @@ def previous_seq_slots_device():
         self.iter_states['num_ctx_requests'] = num_ctx_requests
         self.iter_states['num_ctx_tokens'] = num_ctx_tokens
         self.iter_states['num_generation_tokens'] = num_generation_tokens
-        print(f"DEBUG: is_draft_model: {self.is_draft_model}, inputs: {inputs}")
         return inputs, self.gather_ids_cuda[:len(
             gather_ids)] if self.enable_spec_decode else None
 

tensorrt_llm/_torch/pyexecutor/py_executor.py

@@ -1339,7 +1339,6 @@ def _executor_loop_overlap(self):
                     else:
                         previous_tensors_device = self.previous_batch and self.previous_batch.sample_state and self.previous_batch.sample_state.device
 
-                    print(f"previous_tensors_device: {previous_tensors_device}")
                     batch_outputs = self._forward_step(scheduled_batch,
                                                        previous_tensors_device)
 
@@ -1368,7 +1367,6 @@ def _executor_loop_overlap(self):
                     sample_state = self._sample_async(scheduled_batch,
                                                       batch_outputs)
                     assert sample_state is not None, "Sampling failed"
-                    print(f"sample_state: {sample_state}")
 
                     self._update_request_states(scheduled_batch)