• Docs >
  • Accelerator: IPU training
Shortcuts

Accelerator: IPU training

Audience: IPU users looking to increase performance via mixed precision and analysis tools.

Warning

This is an experimental feature.


Mixed precision & 16 bit precision

Lightning also supports training in mixed precision with IPUs. By default, IPU training will use 32-bit precision. To enable mixed precision, set the precision flag.

Note

Currently there is no dynamic scaling of the loss with mixed precision training.

import lightning.pytorch as pl

model = MyLightningModule()
trainer = pl.Trainer(accelerator="ipu", devices=8, precision=16)
trainer.fit(model)

You can also use pure 16-bit training, where the weights are also in 16-bit precision.

import lightning.pytorch as pl
from lightning.pytorch.strategies import IPUStrategy

model = MyLightningModule()
model = model.half()
trainer = pl.Trainer(accelerator="ipu", devices=8, precision=16)
trainer.fit(model)

PopVision Graph Analyser

PopVision Graph Analyser

Lightning supports integration with the PopVision Graph Analyser Tool. This helps to look at utilization of IPU devices and provides helpful metrics during the lifecycle of your trainer. Once you have gained access, The PopVision Graph Analyser Tool can be downloaded via the GraphCore download website.

Lightning supports dumping all reports to a directory to open using the tool.

import lightning.pytorch as pl
from lightning.pytorch.strategies import IPUStrategy

model = MyLightningModule()
trainer = pl.Trainer(accelerator="ipu", devices=8, strategy=IPUStrategy(autoreport_dir="report_dir/"))
trainer.fit(model)

This will dump all reports to report_dir/ which can then be opened using the Graph Analyser Tool, see Opening Reports.


© Copyright Copyright (c) 2018-2023, Lightning AI et al...

Built with Sphinx using a theme provided by Read the Docs.