tensorflow changelog

Here's the latest scoop on some exciting changes and updates! 🎉 We've been busy improving, fixing, and adding new features to make everything work smoother and faster. Let's dive into the details:

• New Feature: Shared Ownership in tstring 🛠️
  We've added a cool new feature to the tstring class! Now, with tstring::owner<T>, you can enjoy shared ownership of data. This means tstring can manage shared views of data, ensuring it stays alive as long as needed. It's all about safe memory management and keeping things tidy!

• Improvement: Reduction Offloading to YNNPACK 🚀
  We've turbocharged the XLA backend by offloading suitable reductions to YNNPACK. This enhancement checks if a reduction operation fits the bill and, if so, hands it over to the high-performance library for a speed boost. It's all about making things run faster and smoother.

• New Feature: V3 to V2/V1 Conversion 🔄
  Say hello to conversion functions that bridge the gap between different versions in the XLA library. We've also refactored some logic to make it easier to handle reshape dimensions and transpose permutations. This means better interoperability and seamless transitions between versions.

• Improvement: Direct Tensor Reading 📈
  The XLA CPU backend just got a performance boost! We now read directly from input tensors in vectorized reduce operations, eliminating the need for intermediate vectors. This change paves the way for future enhancements like tree-reduction and better bufferization.

• New Feature: Triton Collective Fusion 🔧
  We've introduced a new fusion kind, kTritonCollectiveFusionKind, in the XLA GPU backend. This sets the stage for handling collective operations more efficiently. Actual code emission will follow, but for now, we're laying the groundwork for future improvements.

• New Feature: InprocessSymbolSpecs Serialization 📦
  Serialization just got a whole lot easier with the new KernelSymbolRegistry. It maps unique names to function pointers, making cross-process usage a breeze. This update enhances interoperability and functionality for CUDA and ROCm kernels.

• New Feature: CudaGraph Node Annotations 🖋️
  We've enhanced the trace viewer to show per-node framework names in CudaGraph Nodes. By rewriting annotations, you get more context and clarity in profiling data. It's all about making debugging and profiling a more insightful experience.

• Bugfix: Race Condition in PullTable 🐛
  We squashed a pesky race condition in the PullTable::Handle function. Now, entries are safely checked before erasure, ensuring stability and reliability in concurrent scenarios.

• New Feature: kDotDependent in DimensionInfo 📊
  We've added kDotDependent to DimensionInfo, helping differentiate between gradients and improving scheduling decisions. The HloDimensionAnalysis class now collects this info, enhancing the analysis capabilities of the HLO framework.

• Bugfix: Disable Legacy Emitter Path 🚫
  The legacy emitter path for XLA GPU is now disabled by default. This streamlines the emission process and removes reliance on outdated functionality, making things more efficient.

• Chore: Remove Stride Functions 🧹
  We've cleaned up the XLA types module by removing stride-related functions. This simplifies the codebase and hints at a shift in how strides are handled within the library.

• Bugfix: StableHLO Patch File 🔧
  We've fixed the StableHLO patch file, improving include paths and adding necessary headers. These changes enhance integration and operation within the XLA framework.

That's it for now! Keep an eye out for more updates as we continue to enhance and refine our features. 🚀