fix rl overlap (#7744)

Author: Sunny-bot1

fastdeploy/worker/gpu_model_runner.py

@@ -3043,6 +3043,13 @@ def clear_parameters(self, pid):
         self.clear_cache()
         paddle.device.cuda.empty_cache()
 
+        # clear overlap status
+        self._cached_model_output_data = None
+        self._cached_sampler_output = None
+        self._cached_post_process_event = None
+        self._cached_launch_token_num = -1
+        self._cached_real_bsz = -1
+
         self.dynamic_weight_manager._log_memory("dynamic weight manager clear all memory")
 
     def clear_requests(self):

fastdeploy/worker/input_batch.py

@@ -588,7 +588,6 @@ def reset_share_inputs(self):
             fill_paddle_tensor(self, "step_idx", 0)
             # fill_paddle_tensor(self, "not_need_stop", False)
             fill_paddle_tensor(self, "not_need_stop_device", False)
-            fill_paddle_tensor(self, "sampled_token_ids", -1)
             fill_paddle_tensor(self, "stop_flags", True)
 
             fill_paddle_tensor(self, "bad_tokens", -1)
@@ -726,6 +725,12 @@ def reset_share_inputs(self):
             # Reset other miscellaneous tensors
             fill_paddle_tensor(self, "mask_rollback", 0)
             fill_paddle_tensor(self, "preempted_idx", 0)
+            fill_paddle_tensor(self, "last_preempted_idx", 0)
+
+            # Reset tensors for overlap
+            self.sampled_token_ids = paddle.full([max_num_seqs, 1], -1, dtype="int64").pin_memory()
+            self.seq_lens_this_time_cpu = paddle.full([max_num_seqs, 1], 0, dtype="int32").pin_memory()
+            self.is_block_step_cpu = paddle.full([max_num_seqs], False, dtype="bool").pin_memory()
 
             logger.info("share_inputs reset completed")
         except Exception as e: