torch.js has not been released yet.

torch.nn.BackwardPreHook

export type BackwardPreHook = (
  module: Module,
  gradOutput: (Tensor | null)[]
) => (Tensor | null)[] | void;

Hook called before the backward pass of a module.

BackwardPreHook is executed at the start of a module's backward computation. It receives output gradients (from downstream) and can optionally modify them before they're processed by the module's backward implementation.

Signature:

(module: Module, gradOutput: (Tensor | null)[]) => (Tensor | null)[] | void

Parameters:

module: The module whose backward is about to run
gradOutput: Gradients w.r.t. module outputs (from downstream)

Return Value:

Return modified gradOutput to change what the backward sees
Return void to use original gradients unchanged
Useful for gradient preprocessing or validation

Use Cases:

Inspect incoming gradients before processing
Validate gradient properties before backward
Implement gradient scaling or normalization
Monitor gradient flow at different depths
Implement gradient checkpointing strategies

Examples

const hook: BackwardPreHook = (module, gradOutput) => {
  // Log gradient norm before processing
  console.log('Incoming grad norm:',
    gradOutput[0]?.norm().item() ?? 'null'
  );
  // Don't modify
  return;
};

module.register_full_backward_pre_hook(hook);