DeepSpeedPrecisionPlugin

class pytorch_lightning.plugins.precision.DeepSpeedPrecisionPlugin(precision, amp_type, amp_level=None)

Bases: pytorch_lightning.plugins.precision.precision_plugin.PrecisionPlugin

Precision plugin for DeepSpeed integration.

Parameters

• precision (Union[str, int]) – Double precision (64), full precision (32), half precision (16) or bfloat16 precision (bf16).

• amp_type (str) – The mixed precision backend to use (“native” or “apex”).

• amp_level (Optional[str]) – The optimization level to use (O1, O2, etc…). By default it will be set to “O2” if amp_type is set to “apex”.

Raises

• MisconfigurationException – If using bfloat16 precision and deepspeed<v0.6.

• ValueError – If unsupported precision is provided.

backward(model, closure_loss, optimizer, optimizer_idx, *args, **kwargs)

Performs the actual backpropagation.

Parameters

• model (LightningModule) – the model to be optimized

• closure_loss (Tensor) – the loss value obtained from the closure

• optimizer (Optional[Optimizer]) – current optimizer being used. None if using manual optimization

Return type

None

clip_gradients(optimizer, clip_val=0.0, gradient_clip_algorithm=GradClipAlgorithmType.NORM)

DeepSpeed handles gradient clipping internally.

Return type

None

optimizer_step(model, optimizer, optimizer_idx, closure, **kwargs)

Hook to run the optimizer step.

Return type

Any