scaledDotProductAttention(query:key:value:descriptor:name:)
Creates a scaled dot product attention (SDPA) operation using a descriptor and returns the result tensor.
Declaration
func scaledDotProductAttention(query queryTensor: MPSGraphTensor, key keyTensor: MPSGraphTensor, value valueTensor: MPSGraphTensor, descriptor: MPSGraphSDPADescriptor, name: String?) -> MPSGraphTensorParameters
- queryTensor:
A tensor that represents the query projection.
- keyTensor:
A tensor that represents the key projection.
- valueTensor:
A tensor that represents the value projection.
- descriptor:
A descriptor specifying scale and optional features (mask, isCausal, sinks).
- name:
The name for the operation.
Return Value
A valid MPSGraphTensor object.
Discussion
The descriptor allows configuring an optional attention mask, causal masking, and attention sinks without requiring a separate API method for each combination of features.