IBM Z Deep Learning Compiler 4.2.0

Today we are excited to announce the 4.2.0 update for the IBM Z Deep Learning Compiler (IBM zDLC). The icr.io/ibmz/zdlc:4.2.0 image is available now from the IBM Z and LinuxONE Container Image Registry. An updated tutorial is also available at https://github.com/IBM/zDLC. For those interested for enterprise level support for mission critical workloads, IBM zDLC is included in the AI Toolkit for IBM Z and IBM LinuxONE.

Changes in this release: 

• New compile time options
  • New flags to control placement of individual operators between NNPA and CPU
    • Removed old -execNodesOnCpu flag 
• Updated ONNX-MLIR and ONNX
• ONNX Operators Updates
• Bug fixes and Performance Improvements

New Compile Time Options

--nnpa-save-device-placement-file=<json_file> and --nnpa-load-device-placement-file=<json_file> replaces previous --execNodesOnCpu=<node-name>

The flag to control whether and individual operator goes to the NNPA or the CPU has changed this release. The previous --execNodesOnCpu=<node_name> flag has been removed and replaced by --nnpa-save-device-placement-file=<file> and --nnpa-load-device-placement-file=<file>. These new flags give greater control over device placement, allow saving placement configuration for reuse, and are compatible with models where nodes are not individually named. An example of how to use the new flag is available at https://github.com/IBM/zDLC?tab=readme-ov-file#nnpa-device-placement

Updated to the latest ONNX and ONNX-MLIR

The IBM zDLC is based on on cutting edge open source technology. These updates enable the many features in this release.  In zDLC 4.2.0 we've updated to include the latest released versions of the following:

• ONNX 1.15.0
  
• ONNX-MLIR 0.4.2.0

ONNX Operators Updates

• CPU Operators
  • New
    • Gelu
    • CastLike
    • DFT
    • GroupNormalization
    • LayerNormaliztion
  • Updated for ONNX OpSet 20
    • AveragePool
    • Cast
    • Constant
    • IsInf
    • IsNaN
    • Pad
    • QuantizeLinear
    • Resize
    • Slice
    • Split
• NNPA / Integrated Accelerator for AI Operators
  • GRU added support for optional sequence_lens input

If you have questions on getting started with AI on IBM Z, reach out to us at aionz@us.ibm.com