* Introducing `tf.types.experimental.AtomicFunction` as the fastest way to perform TF computations in Python.

    * Can be accessed through `inference_fn` property of `ConcreteFunction`s
    * Does not support gradients.

        * Before this, the return content was first passed through `jsonable_encoder` to ensure it was a "jsonable" object, like a `dict`, instead of an arbitrary object with attributes (like an ORM model). That's why you should make sure to update your Pydantic models for objects with attributes to use `orm_mode = True`.
        * If you don't have a `response_model`, the return object will still be passed through `jsonable_encoder` first.

"slow path" // for nonces of other lengths. For a 96-bit nonce, the nonce, along // with a four-byte big-endian counter starting at one, is used // directly as the starting counter. For other nonce sizes, the counter // is computed by passing it through the GHASH function. if len(nonce) == gcmStandardNonceSize { copy(counter[:], nonce) counter[gcmBlockSize-1] = 1 } else { lenBlock := make([]byte, 16) byteorder.BEPutUint64(lenBlock[8:], uint64(len(nonce))*8) ghash(counter, H, nonce, lenBlock) } }...

"slow path" // for nonces of other lengths. For a 96-bit nonce, the nonce, along // with a four-byte big-endian counter starting at one, is used // directly as the starting counter. For other nonce sizes, the counter // is computed by passing it through the GHASH function. if len(nonce) == gcmStandardNonceSize { copy(counter[:], nonce) counter[gcmBlockSize-1] = 1 } else { lenBlock := make([]byte, 16) byteorder.BEPutUint64(lenBlock[8:], uint64(len(nonce))*8) ghash(counter, H, nonce, lenBlock) } }...