🧁 Low Precision LayerNorm#
[How to Use] - [Suggested Hyperparameters] - [Technical Details] - [Attribution]
Natural Language Processing,
Low Precision LayerNorm forces
torch.nn.LayerNorm modules to run in float16 or bfloat16 precision, improving utilization. This should not affect final model quality, but in rare cases may cause loss spikes. This algorithm is an alternative to Fused LayerNorm, and should give very similar performance.
How to Use#
# Apply surgery on the model to swap-in the Low Precision LayerNorm using the Composer functional API import composer.functional as cf def training_loop(model, train_loader): cf.apply_low_precision_layernorm(model, precision='amp') opt = torch.optim.Adam(model.parameters()) loss_fn = F.cross_entropy model.train() for X, y in train_loader: y_hat = model(X) loss = loss_fn(y_hat, y) loss.backward() opt.step() opt.zero_grad()
from composer.trainer import Trainer from composer.algorithms import LowPrecisionLayerNorm trainer = Trainer(model=model, train_dataloader=train_dataloader, eval_dataloader=eval_dataloader, max_duration='1ep', algorithms=[LowPrecisionLayerNorm()]) trainer.fit()
Low Precision LayerNorm is implemented by performing model surgery, which looks for instances of
torch.nn.LayerNorm and replaces them with
composer.algorithms.LPLayerNorm. This class is a thin wrapper around
torch.nn.LayerNorm that manually turns autocast off and sets the input dtype to lower precision. In
bf16 mode on PyTorch versions prior to 1.13, Low Precision LayerNorm will fall back to Fused LayerNorm, replacing instances of
Low Precision LayerNorm uses the existing
d_eps from the original model. The functional version of Low Precision LayerNorm allows you to specify the
precision mode, which should be set to the Composer precision format of your model. When using the algorithm through the Composer trainer, Low Precision LayerNorm will use the trainer’s
precision mode automatically.
Low Precision LayerNorm wraps
torch.nn.LayerNorm, forcing the module to run in a lower precision if you have autocast enabled. This depends on the
precision argument passed to Trainer, with
precision='amp_fp16' corresponding to
precision='amp_bf16' corresponding to
This algorithm will have no effect if you are running in
Low Precision LayerNorm is meant to replace our Fused LayerNorm algorithm. The two algorithms achieve very similar throughput. Fused LayerNorm also runs in low precision, but it is a more complex algorithm, since it uses a custom kernel. Since the custom kernel provides no additional speedup, we have replaced it with this simpler algorithm.
✅ Low Precision LayerNorm Improves Training Speed
In our experiments, Low Preicision LayerNorm improves the attainable tradeoffs between training speed and the final quality of the trained model. We recommend using Low Precision LayerNorm.
The Composer implementation of this method and the accompanying documentation were produced by MosaicML.