v0.89.0

What's New

• Circular buffer allocation is now tracked per-core for more accurate L1 pressure analysis, with support for modern core-range formats
• Drastically improved performance on tensor per core allocation rendering and reporting with backwards compatibility support
• Performance table columns can be shown or hidden on demand, with footer totals that stay consistent through column toggles
• Performance table reload is smoother and preserves table state between data refreshes
• Stack trace click-throughs are now available in additional contexts across the visualizer
• Multi-host profiler reports are supported with or without rank suffixes on descriptor files
• Linked operation nodes can be located directly from the operation graph, and duplicate edges between the same endpoints are now collapsed
• Remote sync surfaces clearer errors and falls back to scp when sftp transfer fails

What's Changed

• Improve sync error messaging and add scp fallback by @dcblundell in #1623
• Remove uvicorn dep by @mert-kurttutan in #1627
• Add linked node locate capability by @aidemsined in #1620
• Deduplicate displayed edges by endpoint by @aidemsined in #1621
• Allow users to toggle perf table columns by @dcblundell in #1624
• More gracefully (re)load performance table data by @dcblundell in #1630
• dev: relax deps version by @mert-kurttutan in #1625
• Support per-core CB allocation calculation and core range parsing by @aidemsined in #1628
• Support pre-aggregated buffer-pages chunks for new and legacy formats by @aidemsined in #1631
• Add stack trace click throughs in additional contexts by @dcblundell in #1639
• Support profiler report files with and without rank suffixes by @smountenay-tt in #1642
• Updated performance table footer totals computation by @dcblundell in #1640

New Contributors

• @mert-kurttutan made their first contribution in #1627

Full Changelog: v0.88.0...v0.89.0