(self, config: dict)
| 24 | |
| 25 | class DecoderService(BaseService): |
| 26 | def __init__(self, config: dict): |
| 27 | super().__init__() |
| 28 | self.config = config |
| 29 | self.encoder_engine_rank = int(self.config.get("encoder_engine_rank", 0)) |
| 30 | self.transformer_engine_rank = int(self.config.get("transformer_engine_rank", 1)) |
| 31 | self.decoder_engine_rank = int(self.config.get("decoder_engine_rank", 2)) |
| 32 | self._phase2_rdma_client: Optional[RDMAClient] = None |
| 33 | self._phase2_rdma_buffer: Optional[RDMABuffer] = None |
| 34 | self._centralized_request_mgr = ReqManager() |
| 35 | self._centralized_request_port = REQUEST_POLLING_PORT + self.decoder_engine_rank |
| 36 | data_bootstrap_addr = str(self.config.get("data_bootstrap_addr", "127.0.0.1")) |
| 37 | monitor_bind_host = str(self.config.get("local_hostname", data_bootstrap_addr)) |
| 38 | shared_slots = int(self.config.get("rdma_buffer_slots", "128")) |
| 39 | shared_slot_size = int(self.config.get("rdma_buffer_slot_size", "4096")) |
| 40 | self._phase2_server_ip = str(self.config.get("rdma_phase2_host", data_bootstrap_addr)) |
| 41 | self._phase2_handshake_port = int(self.config.get("rdma_phase2_handshake_port", "5568")) |
| 42 | self._phase2_slots = shared_slots |
| 43 | self._phase2_slot_size = shared_slot_size |
| 44 | self._last_phase2_connect_retry_ts = 0.0 |
| 45 | self.vae_decoder = None |
| 46 | self._rdma_buffers: Dict[int, List[torch.Tensor]] = {} |
| 47 | self.data_mgr = DataManager( |
| 48 | DisaggregationPhase.PHASE2, |
| 49 | DisaggregationMode.DECODE, |
| 50 | ) |
| 51 | self.data_receiver: Dict[int, DataReceiver] = {} |
| 52 | self.req_mgr = ReqManager() |
| 53 | self.reporter = Reporter( |
| 54 | service_type="decoder", |
| 55 | gpu_id=self.decoder_engine_rank, |
| 56 | bind_address=f"tcp://{monitor_bind_host}:{MONITOR_POLLING_PORT + self.decoder_engine_rank}", |
| 57 | ) |
| 58 | self._queue_metrics_lock = threading.Lock() |
| 59 | self._queue_metrics: dict[str, Any] = { |
| 60 | "queue_sizes": {}, |
| 61 | "queue_total_pending": 0, |
| 62 | "all_queues_empty": True, |
| 63 | } |
| 64 | self.reporter.set_extra_metrics_provider(self._get_queue_metrics) |
| 65 | self._reporter_thread: Optional[threading.Thread] = threading.Thread( |
| 66 | target=self.reporter.serve_forever, |
| 67 | name="decoder-reporter", |
| 68 | daemon=True, |
| 69 | ) |
| 70 | self._reporter_thread.start() |
| 71 | self._data_mgr_sidecar = DataMgrSidecar() |
| 72 | self.sync_comm = str(os.getenv("SYNC_COMM", "")).strip().lower() not in ("", "0", "false", "no", "off") |
| 73 | self.load_models() |
| 74 | |
| 75 | def _get_queue_metrics(self) -> dict[str, Any]: |
| 76 | with self._queue_metrics_lock: |
nothing calls this directly
no test coverage detected