Chakra Schema v0.0.4 Release Notes - mlcommons/chakra GitHub Wiki

Introduction

We are excited to announce the release of Chakra schema v0.0.4, dated October 10th, 2023. This release, available in commit version c236cfa, introduces substantial updates and improvements, reflecting our commitment to staying at the forefront of technology and user needs.

Key Changes in v0.0.4

1. Transition to proto3

In an effort to align with the latest technological standards and community practices, Chakra schema has been upgraded from protobuf version 2 (proto2) to version 3 (proto3). This transition is analogous to the widely recognized shift from Python 2 to Python 3, bringing a host of new features and improvements.

2. AttributeProto Message Type Enhancements

2.1. Simplification of Field Declarations

Aligning with proto3’s streamlined approach, we have removed the "required" and "optional" field keywords to simplify the schema and enhance compatibility.

2.2. Expanded Data Type Support

The AttributeProto message type now comprehensively covers all proto3 data types. The introduction of the oneof keyword allows for a more flexible and precise definition of data types, catering to a broader range of use cases.

3. Refined Dependency Encoding

The schema now differentiates more clearly between control dependencies (ctrl_deps) and data dependencies (data_deps). This distinction provides a clearer understanding of the execution flow and data relationships between different schema components.

4. Comprehensive Collective Communication Types

The CollectiveCommType has been extended to include a complete range of collective communication types, ensuring greater versatility in distributed computing scenarios.

5. Introduction of the Message Tensor Type

A new Message Tensor type has been introduced to provide a detailed and structured way to encode tensor information. This enhancement is particularly beneficial for complex data handling and manipulation within the schema.

6. IOInfo for Input/Output Specifications

The introduction of the IOInfo message type offers a more refined and detailed specification for input and output data, promoting greater clarity and efficiency in data handling.

7. Precise Timing Information

To facilitate detailed analysis and optimization of system performance, fields like start_time_micros and duration_micros have been introduced. These fields allow for precise measurement and analysis of operation timings within the schema.

Looking Ahead: Future Revisions

1. Comprehensive Documentation

A detailed written specification will be developed to provide clear guidance on the utilization of new fields, ensuring users can fully leverage the capabilities of the updated schema.

2. Enhanced Concurrency Information

Future revisions will focus on introducing detailed concurrency information for both CPU and GPU, addressing the growing needs for parallel computing efficiency.

3. Revisiting Communicator Group Information

The handling of communicator group information, particularly in relation to the inputs field, is slated for a comprehensive review and potential restructuring.

Conclusion

With these updates, Chakra schema v0.0.4 takes a significant step forward in providing a more powerful, flexible, and user-friendly tool for our community. We are committed to continuous improvement and eagerly anticipate your feedback and contributions.