Shortcuts

DeepSpeedPrecisionPlugin

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

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(tensor, model, optimizer, optimizer_idx, *args, **kwargs)[source]

Performs back-propagation using DeepSpeed’s engine.

Parameters:
  • tensor (Tensor) – the loss tensor

  • model (LightningModule) – the model to be optimized

  • optimizer (Optional[Steppable]) – ignored for DeepSpeed

  • optimizer_idx (Optional[int]) – ignored for DeepSpeed

  • *args (Any) – additional positional arguments for the deepspeed.DeepSpeedEngine.backward() call

  • **kwargs (Any) – additional keyword arguments for the deepspeed.DeepSpeedEngine.backward() call

Return type:

None

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

DeepSpeed handles gradient clipping internally.

Return type:

None

optimizer_step(optimizer, model, optimizer_idx, closure, **kwargs)[source]

Hook to run the optimizer step.

Return type:

Any