Initialize device and construct model runner
(self)
| 56 | pass |
| 57 | |
| 58 | def init_device(self): |
| 59 | """ |
| 60 | Initialize device and construct model runner |
| 61 | """ |
| 62 | self.max_chips_per_node = 16 if current_platform.is_iluvatar() else 8 |
| 63 | if self.device_config.device_type == "cuda" and paddle.device.is_compiled_with_cuda(): |
| 64 | # Set environment variable |
| 65 | self.device_ids = self.parallel_config.device_ids.split(",") |
| 66 | self.device = f"gpu:{self.local_rank % self.max_chips_per_node}" |
| 67 | paddle.device.set_device(self.device) |
| 68 | paddle.set_default_dtype(self.model_config.dtype) |
| 69 | |
| 70 | gc.collect() |
| 71 | paddle.device.cuda.empty_cache() |
| 72 | if ( |
| 73 | not self.parallel_config.disable_custom_all_reduce |
| 74 | and self.parallel_config.tensor_parallel_size > 1 |
| 75 | and paddle.is_compiled_with_cuda() |
| 76 | ): |
| 77 | from fastdeploy.distributed.communication import use_custom_allreduce |
| 78 | |
| 79 | use_custom_allreduce(self.fd_config.parallel_config.tp_group) |
| 80 | else: |
| 81 | raise RuntimeError(f"Not support device type: {self.device_config.device}") |
| 82 | |
| 83 | if self.local_rank == 0: |
| 84 | report_usage_stats(self.fd_config) |
| 85 | |
| 86 | set_random_seed(self.fd_config.model_config.seed) |
| 87 | # Construct model runner |
| 88 | self.model_runner: ModelRunnerBase = ModelRunner( |
| 89 | fd_config=self.fd_config, |
| 90 | device=self.device, |
| 91 | device_id=int(self.device_ids[self.local_rank % self.max_chips_per_node]), |
| 92 | rank=self.rank, |
| 93 | local_rank=self.local_rank, |
| 94 | ) |
| 95 | |
| 96 | def exist_prefill(self): |
| 97 | """ |
nothing calls this directly
no test coverage detected