Introduction Type Safety Tensor Expressions Tensor Indexing Einsum Einops Autograd Training a Model Profiling & Memory PyTorch Migration Best Practices Runtimes Performance PyTorch Compatibility Benchmarks DType Coverage

torch.BinaryOpConfig

export interface BinaryOpConfig<Op extends BinaryOpNames> {
  /**
   * Operation name (must match OpSchemas).
   */
  readonly op: Op;

  /**
   * CPU kernel function. Receives two values and returns the result.
   * For float32 dtype.
   */
  readonly cpu: (a: number, b: number) => number;

  /**
   * Optional CPU kernel for int32 dtype.
   * If not provided and int32 is in dtypes, uses the float32 kernel.
   */
  readonly cpuInt32?: (a: number, b: number) => number;

  /**
   * WebGPU shader entry point name (e.g., 'add_op', 'mul_op').
   */
  readonly shaderEntry: string;

  /**
   * Dtypes to register. Defaults to ['float32', 'int32'].
   */
  readonly dtypes?: readonly BinaryDType[];

  /**
   * Backward function for autograd.
   * If not provided, the operation is not differentiable.
   */
  readonly backward?: BinaryBackwardFn;

  /**
   * Setup context function for backward.
   * Defaults to saving both inputs.
   */
  readonly setupContext?: (
    ctx: GradContext,
    inputs: { a: TensorLike; b: TensorLike },
    output: TensorLike
  ) => void;

  /**
   * What to save for backward pass.
   * 'inputs' - saves both input tensors (default)
   * 'output' - saves the output tensor
   * 'none' - saves nothing
   * 'a' - saves only the first input
   * 'b' - saves only the second input
   */
  readonly saveForBackward?: 'inputs' | 'output' | 'none' | 'a' | 'b';

  /**
   * Tensor method name to call. Defaults to the op name.
   */
  readonly tensorMethod?: string;
}

Opextends BinaryOpNames

readonlyop(Op): – Operation name (must match OpSchemas).
readonlycpu((a: number, b: number) => number): – CPU kernel function. Receives two values and returns the result. For float32 dtype.
readonlycpuInt32((a: number, b: number) => number)optional: – Optional CPU kernel for int32 dtype. If not provided and int32 is in dtypes, uses the float32 kernel.
readonlyshaderEntry(string): – WebGPU shader entry point name (e.g., 'add_op', 'mul_op').
readonlydtypes(readonly BinaryDType[])optional: – Dtypes to register. Defaults to ['float32', 'int32'].
readonlybackward(BinaryBackwardFn)optional: – Backward function for autograd. If not provided, the operation is not differentiable.
readonlysetupContext(( ctx: GradContext, inputs: { a: TensorLike; b: TensorLike }, output: TensorLike ) => void)optional: – Setup context function for backward. Defaults to saving both inputs.
readonlysaveForBackward('inputs' | 'output' | 'none' | 'a' | 'b')optional: – What to save for backward pass. 'inputs' - saves both input tensors (default) 'output' - saves the output tensor 'none' - saves nothing 'a' - saves only the first input 'b' - saves only the second input
readonlytensorMethod(string)optional: – Tensor method name to call. Defaults to the op name.

Configuration for registering a binary operation.