v0.92.0

What's New

• MLIR viewer gets a floating op-name filter — live substring match with dim/highlight, +N badges on collapsed anchors that own buried matches, prev/next navigation with viewport pan, Cmd/Ctrl+F to focus, Esc to clear
• MLIR uploads now accept multiple files in one go, showing per-file processing status and letting the user pick which graph to activate from a results overlay
• The Performance Table now includes per-op kernel timing, and its cell colours were aligned with tt-perf-report so the two tools read the same
• Operation list rows now show an inline kernel-duration bar, matching the perf-overlay colour scale so hot ops are visible without opening the Buffer Summary
• Cluster view supports multi-host topologies: reads mesh coordinates from the new descriptor format, ships an original ↔ condensed layout toggle for multi-host reports, and routes inter-host ETH links along the free sides of each chip to reduce line crossings
• Globally allocated CBs in the per-core L1 stack now render as translucent tinted fills with a solid border, and the memory legend tooltips report the alias, dimensions, and dtype so the user learns something from the hover
• Docs site gets OpenSearch indexing, a sidebar search, and SPDX headers across the tree
• CI workflows now run in parallel where possible, and the shared workflow actions were upgraded
• Git repository url and current commit hash added to database
• The TTNN comparison mode feature now writes its records through ttnn.graph_report (via a graph_capture.comparison_records.json sidecar), retiring the legacy database.py path so there's a single source of truth for report-DB creation
• Mesh descriptor YAML files are now copied verbatim from generated/fabric into TTNN reports instead of being concatenated with -- separators, so multi-host cluster topology can consume them as-is
• Fixed a graph processor crash on the first non-host op in multi-host runs, so remote-rank ops now land in the graph capture JSON and report DB (ttnn.all_gather / CCL ops still pending a follow-up)

What's Changed

• Align performance table cell colours with tt-perf-report by @dcblundell in #1686
• Multi-host cluster topology rendering (#1510) by @aidemsined in #1677
• Render globally-allocated CBs as outlined overlays in the L1 stack by @aidemsined in #1687
• Cherry pick doc updates for dev by @dcblundell in #1701
• Upgrade shared workflow actions by @dcblundell in #1704
• feat(docs): add OpenSearch indexing, sidebar search, and SPDX headers by @ctr-fmamedov-crypto in #1702
• Use parallel execution in workflows by @dcblundell in #1694
• Add per-op kernel timing to Performance Table by @dcblundell in #1689
• Support multiple MLIR file uploads and processing by @dcblundell in #1693
• Add inline kernel-duration bars to operation list rows by @aidemsined in #1695
• Add MLIR op-name filter with dim / navigation / buried-match surfacing by @aidemsined in #1708
• MLIR: Refactor match navigation and filter handling by @aidemsined in #1709

Full Changelog: v0.91.0...v0.92.0