Lightning-AI/pytorch-lightning

Make lazy initialization in plugins more robust

Open

#7,650 建立於 2021年5月21日

在 GitHub 查看
 (4 留言) (0 反應) (1 負責人)Python (3,233 fork)batch import
distributedfeaturegood first issuehelp wantedlet's do it!

倉庫指標

Star
 (26,687 star)
PR 合併指標
 (平均合併 9天 15小時) (30 天內合併 3 個 PR)

描述

🚀 Feature

We currently set some attributes like sync_batch_norm, num_nodes, devices etc. lazily in the accelerator connector, so the user does not have to provide them at instantiation.

Example:

Trainer(gpus=4, plugins=DDPPlugin(find_unused_parameters=True))  # plugin may require gpus, num_nodes etc.

# AcceleratorConnector does this:
training_type_plugin.num_nodes = ...

This is fragile. Some attributes may have to be recomputed based on the order in which others are set.

Pitch

Provide one single lazy init method that takes all arguments required. The plugin is responsible for making sure dependencies are resolved in one place:

class DDPPlugin(...):

    def lazy_init(self, **kwargs):
         self.num_nodes = kwargs.get("num_nodes")
         self.num_processes = ...

With lazy_init, the _configure_launcher method (#11643) would become obsolete. It can be merged together into lazy_init.

貢獻者指南