(self)
| 1076 | self.accept_token_num_per_head[i] += 1 |
| 1077 | |
| 1078 | def clear_data(self): |
| 1079 | if envs.ENABLE_V1_KVCACHE_SCHEDULER: |
| 1080 | self.resource_manager.clear_data() |
| 1081 | for i in range(self.resource_manager.max_num_seqs): |
| 1082 | if self.resource_manager.stop_flags[i]: |
| 1083 | continue |
| 1084 | task = self.resource_manager.tasks_list[i] |
| 1085 | result = RequestOutput( |
| 1086 | request_id=task.request_id, |
| 1087 | outputs=CompletionOutput( |
| 1088 | index=i, |
| 1089 | send_idx=self.tokens_counter[task.request_id], |
| 1090 | token_ids=task.eos_token_ids, |
| 1091 | draft_token_ids=[], |
| 1092 | ), |
| 1093 | finished=True, |
| 1094 | metrics=RequestMetrics( |
| 1095 | arrival_time=time.time(), |
| 1096 | request_start_time=task.metrics.arrival_time, |
| 1097 | ), |
| 1098 | ) |
| 1099 | is_prefill = task.disaggregate_info is not None and task.disaggregate_info["role"] == "prefill" |
| 1100 | self._recycle_resources(task.request_id, i, task, result, is_prefill) |
| 1101 | llm_logger.warning(f"clear data for task {task.request_id}") |
| 1102 | |
| 1103 | |
| 1104 | class WarmUpTokenProcessor(TokenProcessor): |
nothing calls this directly
no test coverage detected