and works in the DTensor training use case.
    *   Improved masking support for
        [`tf.keras.layers.MultiHeadAttention`](https://www.tensorflow.org/api_docs/python/tf/keras/layers/MultiHeadAttention).
        *   Implicit masks for `query`, `key` and `value` inputs will
            automatically be used to compute a correct attention mask for the
            layer. These padding masks will be combined with any

0/subtle" "errors" ) // A DecapsulationKey1024 is the secret key used to decapsulate a shared key from a // ciphertext. It includes various precomputed values. type DecapsulationKey1024 struct { d [32]byte // decapsulation key seed z [32]byte // implicit rejection sampling seed ρ [32]byte // sampleNTT seed for A, stored for the encapsulation key h [32]byte // H(ek), stored for ML-KEM.Decaps_internal encryptionKey1024 decryptionKey1024 } // Bytes returns the decapsulation key as a 64-byte seed in...