*   `tf.nn.embedding_lookup_sparse`

    *   Optimized this function for some cases by fusing internal operations.

*   `tf.saved_model.SaveOptions`

* Improved **correctness** in corner cases.
* **Safer** types.
* Better **performance** and **less energy** consumption.
* Better **extensibility**.
* etc.

...all this while keeping the **same Python API**. In most of the cases, for simple models, you can simply upgrade the Pydantic version and get all the benefits. 🚀

subtraction tells us if // x >= m as a side effect, and taking care of the broken invariant is // highly undesirable (see https://go.dev/issue/13907). copy(x.reset(n).limbs, T[n:]) x.maybeSubtractModulus(choice(c), m) // The following specialized cases follow the exact same algorithm, but // optimized for the sizes most used in RSA. addMulVVW is implemented in // assembly with loop unrolling depending on the architecture and bounds // checks are removed by the compiler thanks to the constant size....

subtraction tells us if // x >= m as a side effect, and taking care of the broken invariant is // highly undesirable (see https://go.dev/issue/13907). copy(x.reset(n).limbs, T[n:]) x.maybeSubtractModulus(choice(c), m) // The following specialized cases follow the exact same algorithm, but // optimized for the sizes most used in RSA. addMulVVW is implemented in // assembly with loop unrolling depending on the architecture and bounds // checks are removed by the compiler thanks to the constant size....