Skip to content

v0.91.0

Latest

Choose a tag to compare

@aidemsined aidemsined released this 24 Jun 19:23
· 25 commits to dev since this release

What's New

  • The Buffer Summary timeline now highlights the top-N slowest ops with a colour-coded rank badge in the y-axis gutter and a clickable rail to the right, ranked by kernel duration, op-to-op gap, DRAM utilisation, FLOPS utilisation, or L1 fullness
  • MLIR upload UX is significantly cleaner — the global progress overlay covers the full upload → processing → ready flow with clearer messaging, and unreachable-server errors now point the user at MLIR specifically instead of a generic networking failure
  • The cluster view rendering was refined for clarity (port placement, link routing, and host layout heuristics)
  • A "show/hide globally allocated CBs" toggle was added to the Circular Buffer Pressure modal so dense ops can be read without aliased CBs crowding the per-core grid; peak and total counts stay anchored to the pressure-relevant subset
  • The per-DeviceOp memory legend rows no longer overlap the right-side CBs / Buffers / Total summary panel, and the buffer-allocate / buffer-deallocate / CB rows now align under a single legend grid
  • Switching between reports no longer surfaces stale operation-detail data from the previously open report — caches are now keyed by the active report path
  • Successful memory ↔ performance report link history is now remembered and surfaced as a badge in the report selection list, prioritising previously linked counterparts (experimental, dev-flag gated)
  • Contributors now use UV across all Python tooling, the lockfile is kept in sync on release, and CI enforces lockfile consistency viauv lock --check
  • Fixed a bug in TTNN graph processor that prevented non-host ops from being recorded to the TTNN graph capture JSON file and report db in multihost
link topology mlir 611291382-1d75dcff-6922-48dd-8a1b-b2d6f9007fb6 610024416-69455ba0-89f8-435a-82f0-94cc2195d508

What's Changed

Full Changelog: v0.90.0...v0.91.0