*   CUDA 9.0 and cuDNN 7 support.
*   Accelerated Linear Algebra (XLA):
    *   Add `complex64` support to XLA compiler.
    *   `bfloat` support is now added to XLA infrastructure.
    *   Make `ClusterSpec` propagation work with XLA devices.
    *   Use a deterministic executor to generate XLA graph.
*   `tf.contrib`:
    *   `tf.contrib.distributions`:
    *   Add `tf.contrib.distributions.Autoregressive`.

## 0.122.1 (2025-11-30)

### Fixes

* 🐛 Fix hierarchical security scope propagation. PR [#5624](https://github.com/fastapi/fastapi/pull/5624) by [@kristjanvalur](https://github.com/kristjanvalur).

### Docs

b.l3 v.l4 = a.l4 + b.l4 // Using the generic implementation here is actually faster than the // assembly. Probably because the body of this function is so simple that // the compiler can figure out better optimizations by inlining the carry // propagation. return v.carryPropagateGeneri() } // Subtract sets v = a - b, and returns v. func (v *Element) Subtract(a, b *Element) *Element { // We first add 2 * p, to guarantee the subtraction won't underflow, and // then subtract b (which can be up to 2^255...

v.l0 = x0lo + 19*x4hi // carried over per the reduction identity v.l1 = x1lo + x0hi v.l2 = x2lo + x1hi v.l3 = x3lo + x2hi v.l4 = x4lo + x3hi // The hi portions are going to be only 32 bits, plus any previous excess, // so we can skip the carry propagation. return v } // mul51 returns lo + hi * 2⁵¹ = a * b. func mul51(a uint64, b uint32) (lo uint64, hi uint64) { mh, ml := bits.Mul64(a, uint64(b)) lo = ml & maskLow51Bits hi = (mh << 13) | (ml >> 51) return } // Pow22523 set v = x^((p-5)/8), and returns...