配置环境变量
(self)
| 457 | p.cleanup() |
| 458 | |
| 459 | def _setting_environ_variables(self): |
| 460 | """ |
| 461 | 配置环境变量 |
| 462 | """ |
| 463 | variables = { |
| 464 | "ENABLE_FASTDEPLOY_LOAD_MODEL_CONCURRENCY": 0, |
| 465 | "LOAD_STATE_DICT_THREAD_NUM": len(self.cfg.parallel_config.device_ids.split(",")), |
| 466 | "PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION": "python", |
| 467 | "NCCL_ALGO": "Ring", |
| 468 | "FLAGS_max_partition_size": int(os.getenv("FLAGS_max_partition_size", 1024)), |
| 469 | "OMP_NUM_THREADS": 3, |
| 470 | "FD_ENABLE_PDL": envs.FD_ENABLE_PDL, |
| 471 | } |
| 472 | # environment variables needed by Dy2St |
| 473 | variables.update( |
| 474 | { |
| 475 | "SOT_LOG_LEVEL": os.getenv("SOT_LOG_LEVEL", default="0"), |
| 476 | "SOT_UNSAFE_CACHE_FASTPATH": os.getenv("SOT_UNSAFE_CACHE_FASTPATH", default="1"), |
| 477 | "SOT_ENABLE_0_SIZE_FALLBACK": os.getenv("SOT_ENABLE_0_SIZE_FALLBACK", default="0"), |
| 478 | "SOT_SPECIALIZED_DIM_NUMBERS": os.getenv("SOT_SPECIALIZED_DIM_NUMBERS", default="no"), |
| 479 | "SOT_ENABLE_COMPILE_TIME_LIMIT": os.getenv("SOT_ENABLE_COMPILE_TIME_LIMIT", default="0"), |
| 480 | "FLAGS_specialize_device_in_dy2st": os.getenv("FLAGS_specialize_device_in_dy2st", default="1"), |
| 481 | "FLAGS_enable_async_fast_gc": os.getenv("FLAGS_enable_async_fast_gc", default="0"), |
| 482 | "FLAGS_pir_interpreter_record_stream_for_gc_cache": os.getenv( |
| 483 | "FLAGS_pir_interpreter_record_stream_for_gc_cache", default="1" |
| 484 | ), |
| 485 | "FLAGS_parameters_persistent_mode_in_dy2st": os.getenv( |
| 486 | "FLAGS_parameters_persistent_mode_in_dy2st", default="1" |
| 487 | ), |
| 488 | } |
| 489 | ) |
| 490 | |
| 491 | if self.cfg.scheduler_config.splitwise_role != "mixed": |
| 492 | if envs.ENABLE_V1_KVCACHE_SCHEDULER: |
| 493 | variables["FLAGS_use_pd_disaggregation_per_chunk"] = 1 |
| 494 | else: |
| 495 | variables["FLAGS_use_pd_disaggregation"] = 1 |
| 496 | # TODO dynamic load environment variable |
| 497 | if self.cfg.scheduler_config.splitwise_role == "prefill": |
| 498 | variables["FLAGS_fmt_write_cache_completed_signal"] = 1 |
| 499 | |
| 500 | command_prefix = "" |
| 501 | for k, v in variables.items(): |
| 502 | command_prefix += f"{k}={v} " |
| 503 | return command_prefix |
| 504 | |
| 505 | def _start_worker_service(self): |
| 506 | """ |