| 28 | |
| 29 | |
| 30 | class DcuWorker(GpuWorker): |
| 31 | """ """ |
| 32 | |
| 33 | def __init__( |
| 34 | self, |
| 35 | fd_config: FDConfig, |
| 36 | local_rank: int, |
| 37 | rank: int, |
| 38 | ): |
| 39 | super().__init__( |
| 40 | fd_config=fd_config, |
| 41 | local_rank=local_rank, |
| 42 | rank=rank, |
| 43 | ) |
| 44 | pass |
| 45 | |
| 46 | def init_device(self): |
| 47 | """ |
| 48 | Initialize device and construct model runner |
| 49 | """ |
| 50 | self.max_chips_per_node = 8 |
| 51 | if self.device_config.device_type == "cuda" and paddle.device.is_compiled_with_cuda(): |
| 52 | # Set environment variable |
| 53 | self.device_ids = self.parallel_config.device_ids.split(",") |
| 54 | self.device = f"gpu:{self.local_rank % self.max_chips_per_node}" |
| 55 | paddle.device.set_device(self.device) |
| 56 | paddle.set_default_dtype(self.model_config.dtype) |
| 57 | |
| 58 | gc.collect() |
| 59 | paddle.device.cuda.empty_cache() |
| 60 | if ( |
| 61 | self.parallel_config.enable_custom_all_reduce |
| 62 | and self.parallel_config.tensor_parallel_size > 1 |
| 63 | and paddle.is_compiled_with_cuda() |
| 64 | ): |
| 65 | from fastdeploy.distributed.communication import use_custom_allreduce |
| 66 | |
| 67 | use_custom_allreduce(self.fd_config.parallel_config.tp_group) |
| 68 | else: |
| 69 | raise RuntimeError(f"Not support device type: {self.device_config.device}") |
| 70 | |
| 71 | set_random_seed(self.fd_config.model_config.seed) |
| 72 | # Construct model runner |
| 73 | self.model_runner: DCUModelRunner = DCUModelRunner( |
| 74 | fd_config=self.fd_config, |
| 75 | device=self.device, |
| 76 | device_id=self.device_ids[self.local_rank % self.max_chips_per_node], |
| 77 | rank=self.rank, |
| 78 | local_rank=self.local_rank, |
| 79 | ) |
| 80 | |
| 81 | def determine_available_memory(self) -> int: |
| 82 | """ |
| 83 | Profiles the peak memory usage of the model to determine how much |
| 84 | memory can be used for KV cache without OOMs. |
| 85 | |
| 86 | The engine will first conduct a profiling of the existing memory usage. |
| 87 | Then, it calculate the maximum possible number of GPU and CPU blocks |