v0.90.0
What's New
- MLIR end-to-end very early beta integration with tt-mlir's model explorer is available as a beta feature from the home route (please refer to tt-mlir for how to get going). Did we mention its in beta?
- Circular buffer allocations are now visualizable per core through a new pressure modal, including a zoomed view of the selected core's address window
- Globally allocated circular buffers are rendered as outline-only tensor views so they no longer double-count against per-core L1 pressure or peak load
- Condensed circular buffer envelope now uses the lowest allocated address as its lower bound, so the summary stripe accurately covers what was actually allocated
- Selecting a different node in the MLIR view now reliably clears any pending focus, fixing the stale-focus behavior on the initial graph load
- Tensor sharding visualizations no longer mutate cached buffer chunk data, preventing stale state across re-renders
- Python dependency management has moved to UV for easier local setup
- The pinned UV version was relaxed to a minimum bound so contributors can use any recent UV release
- Local
.pnpm-store/cache directory is now gitignored to keep workspaces clean - ttnn.graph_report tool was updated to merge graph capture JSON files from multiple hosts into a single report database, and track rank throughout the report
What's Changed
- Fix calculateCondensed to use min(address) as the envelope lower bound by @aidemsined in #1654
- Use UV for Python dependency version management by @dcblundell in #1645
- MLIR: clear pendingFocusNodeIdRef when selection changes by @aidemsined in #1649
- Relax pinned uv version to a minimum bound by @smountenay-tt in #1658
- E2E integration with tt-mlir/model explorer by @dcblundell in #1647
- chore: ignore local .pnpm-store/ by @aidemsined in #1660
- Show circular buffer allocation view per core by @aidemsined in #1648
- Stop mutating cached BufferChunk objects in TensorVisualisationComponent by @aidemsined in #1650
- Treat globally_allocated CBs as tensor views, not separate allocations by @aidemsined in #1656
- Adds beta tag to MLIR on home route by @dcblundell in #1661
Full Changelog: v0.89.0...v0.90.0