Release Highlights - Release Highlights - 57300

ZenDNN User Guide (57300)
Document ID
57300
Release Date
2026-04-13
Revision
5.2.1 English

This release of AMD's CPU solution for TensorFlow provides a binary built with the PluggableDevice approach.

zentf v5.2.1 Release Highlights

  • TF 2.21 Integration
    • zentf 5.2.1 is built for and validated against TensorFlow v2.21.0.
    • Upgraded from Bazel 7.4.1 to 7.7.0.
    • Python 3.10 - 3.13 -- Compatibility with TensorFlow (support for Python 3.9 discontinued).
    • As TensorFlow-Java has not been released for TensorFlow 2.21.0, zentf supports TensorFlow-Java main(75402bef) via source build only.
  • Backward Build Compatibility Support: TF 2.16 to 2.21
    • zentf can now be built from source against TensorFlow versions 2.16.0 through 2.21.0 from a single unified codebase.
    • The ./configure script auto-detects the installed TensorFlow version and applies the matching build configuration — no manual intervention required.
    • Per-version build configurations (workspace, WORKSPACE, build_config.bzl, BUILD.tpl) are maintained under version_configs/ for TF 2.19, 2.20, and 2.21. Note that TF 2.16–2.18 share the TF 2.19 configuration.
    • Bazel version and third-party dependency settings (protobuf, abseil, rules_cc) are automatically adjusted to match the target TensorFlow version.
    • The standard build workflow (./configure > bazel build) remains unchanged; version adaptation is fully transparent.

zentf v5.2 Release Highlights

  • TF 2.20 Migration
    • zentf 5.2.0 is built for and validated against TensorFlow v2.20.0.
    • Bazel 7.4.1: Upgraded from Bazel 5.3-6.5 range to a single supported version (7.4.1).
    • Python 3.9 - 3.13: Extended Python version support to include Python 3.13.
    • As TF JAVA is not released with 2.20 version, zentf is supported with main (75402bef) branch from TensoFlow-Java through source build only.
  • Migration from legacy ZenDNN library to ZenDNNL
    • CMake-based ZenDNNL integration using rules_foreign_cc.
    • All operator kernels (MatMul, Conv2D, BatchMatMul, Softmax, Pooling) have been rewritten to use the ZenDNNL Low Overhead API (LOA), replacing the legacy ZenDNN primitives.
    • Old third-party dependencies on zen_dnn and amd_blis (BLIS) have been removed and replaced with ZenDNNL with integrated AOCL-DLP.
  • Removed Legacy Components
    • Mempool optimization has been completely removed and equivalent performance has been achieved using jemalloc as the memory allocator instead.
    • INT8 support has been removed.
    • Removal of non-performant ops: ZenTranspose, ZenReshape, Binary ops.