Release v0.5.0

Summary

Added

• Added support for AMD GPUs.
• Update pyproject.toml to workaround missing stub packages for yaml.
• Add trace format validator
• Added multiple trace filter classes and demos.
• Added enhanced trace call stack graph implementation.
• Added memory timeline view.
• Added support for trace parser customization.
• Added support for H100 traces.
• Add nccl collective fields to parser config
• Queue length analysis: Add feature to compute time blocked on a stream hitting max queue length.
• Add kernel_backend to parser config for Triton / torch.compile() support.
• Add analyses features for GPU user annotation attribution at trace and kernel level.
• Add support to parse all trace event args.

New Feature: Critical Path Analysis

• Added lightweight critical path analysis feature.
• Critical path analysis features: event attribution and summary()
• Critical path analysis fixes: fixing async memcpy and adding GPU to CPU event based synchronization.
• Added save and restore feature for critical path graph.
• Added save and restore feature for critical path graph.
• Fixes bug in Critical path analysis relating to listing out the edges on the critical path.
• Updated critical path analysis with edge attribution.
• Improvement: allow filtering of flow events in the overlaid trace.

Changed

• Change test data path in unittests from relative path to real path to support running test within IDEs.
• Add a workaround for overlapping events when using ns resolution traces (pytorch/pytorch#122425)
• Better handling of CUDA sync evaents with steam = -1
• Fix ijson metadata parser for some corner cases
• Add an option for ns rounding and cover ijson loading with it.
• Updated Trace() api to specify a list of files and auto figure out ranks.

Fixed

• Fixed issue #65 to handle floating point counter values in cupti_counter_analysis.

Full Changelog: v0.2.0...v0.5.0