|
| 1 | +from typing import Any, Type |
| 2 | + |
| 3 | +from lightning_app import structures |
| 4 | +from lightning_app.core.flow import LightningFlow |
| 5 | +from lightning_app.core.work import LightningWork |
| 6 | +from lightning_app.utilities.enum import WorkStageStatus |
| 7 | +from lightning_app.utilities.packaging.cloud_compute import CloudCompute |
| 8 | + |
| 9 | + |
| 10 | +class MultiNode(LightningFlow): |
| 11 | + def __init__( |
| 12 | + self, |
| 13 | + work_cls: Type["LightningWork"], |
| 14 | + nodes: int, |
| 15 | + cloud_compute: "CloudCompute", |
| 16 | + *work_args: Any, |
| 17 | + **work_kwargs: Any, |
| 18 | + ) -> None: |
| 19 | + """This component enables performing distributed multi-node multi-device training. |
| 20 | +
|
| 21 | + Example:: |
| 22 | +
|
| 23 | + import torch |
| 24 | +
|
| 25 | + import lightning as L |
| 26 | + from lightning.components import MultiNode |
| 27 | +
|
| 28 | + class AnyDistributedComponent(L.LightningWork): |
| 29 | + def run( |
| 30 | + self, |
| 31 | + main_address: str, |
| 32 | + main_port: int, |
| 33 | + node_rank: int, |
| 34 | + ): |
| 35 | + print(f"ADD YOUR DISTRIBUTED CODE: {main_address} {main_port} {node_rank}") |
| 36 | +
|
| 37 | +
|
| 38 | + compute = L.CloudCompute("gpu") |
| 39 | + app = L.LightningApp( |
| 40 | + MultiNode( |
| 41 | + AnyDistributedComponent, |
| 42 | + nodes=8, |
| 43 | + cloud_compute=compute, |
| 44 | + ) |
| 45 | + ) |
| 46 | +
|
| 47 | + Arguments: |
| 48 | + work_cls: The work to be executed |
| 49 | + nodes: Number of nodes. |
| 50 | + cloud_compute: The cloud compute object used in the cloud. |
| 51 | + work_args: Arguments to be provided to the work on instantiation. |
| 52 | + work_kwargs: Keywords arguments to be provided to the work on instantiation. |
| 53 | + """ |
| 54 | + super().__init__() |
| 55 | + self.ws = structures.List() |
| 56 | + self._work_cls = work_cls |
| 57 | + self.nodes = nodes |
| 58 | + self._cloud_compute = cloud_compute |
| 59 | + self._work_args = work_args |
| 60 | + self._work_kwargs = work_kwargs |
| 61 | + self.has_initialized = False |
| 62 | + |
| 63 | + def run(self) -> None: |
| 64 | + # 1. Create & start the works |
| 65 | + if not self.has_initialized: |
| 66 | + for node_rank in range(self.nodes): |
| 67 | + self.ws.append( |
| 68 | + self._work_cls( |
| 69 | + *self._work_args, |
| 70 | + cloud_compute=self._cloud_compute, |
| 71 | + **self._work_kwargs, |
| 72 | + parallel=True, |
| 73 | + ) |
| 74 | + ) |
| 75 | + # Starting node `node_rank`` ... |
| 76 | + self.ws[-1].start() |
| 77 | + self.has_initialized = True |
| 78 | + |
| 79 | + # 2. Wait for all machines to be started ! |
| 80 | + if all(w.status.stage == WorkStageStatus.STARTED for w in self.ws): |
| 81 | + return |
| 82 | + |
| 83 | + # Loop over all node machines |
| 84 | + for node_rank in range(self.nodes): |
| 85 | + |
| 86 | + # 3. Run the user code in a distributed way ! |
| 87 | + self.ws[node_rank].run( |
| 88 | + main_address=self.ws[0].internal_ip, |
| 89 | + main_port=self.ws[0].port, |
| 90 | + node_rank=node_rank, |
| 91 | + ) |
| 92 | + |
| 93 | + # 4. Stop the machine when finished. |
| 94 | + if self.ws[node_rank].has_succeeded: |
| 95 | + self.ws[node_rank].stop() |
0 commit comments