Highlights

• 🚀 dace:cpu_KJI backend with Fortran friendly memory layout
• 🚀 LocalState: a State-like class for organizing locals (comes with performance benefits)
• 🚀 Optimized C-call overhead
• ✨ New package ndsl.xumpy: Performance portable API for writing non-stencil numerics in python
• 🌎 A couple more basic stencils
• 🐞 Fix index returned by column min/max functions
• 🧪 Experimental work on axis merge, cache optimizations and automatic code labeling

⚠️ Breaking changes

With this version of NDSL, we introduced the following breaking changes.

• NDSLRuntime requires StencilFactory instead of previous config
• ZarrMonitor now requires a Comm object in the constructor
• Removed deprecated CopyCornersXY
• Removed get_tile_number() - use get_tile_index() instead
• Removed deprecated QuantityFactory.from_backend() - use the class constructor instead
• Removed deprecated quantity.gt4py_backend - use quantity.backend instead
• Removed deprecated NullComm class - use LocalComm instead

Deprecations

This version of NDSL added the following deprecation warnings:

• SubtileGridSizer will require a backend to initialize
• Duplicate basic stencil definitions are deprecated, e.g. copy_defn(...) has to be replaced by copy(...)

What's Changed

• [API Update] NDSLRuntime uses the wider StencilFactory rather than a narrow config by @FlorianDeconinck in #335
• 🔧 Harden Quantity to hybrid usage by @FlorianDeconinck in #336
• 🌎 A few more basic stencils by @FlorianDeconinck in #337
• 🚀 LocalState class: a state-like class for organizing Locals by @FlorianDeconinck in #338
• [Fix | ScheduleTreeOpt | Experimental] CartesianAxisMerge correct parenting and RefineTransient correct logic by @FlorianDeconinck in #341
• [Feature] LocalState allocation as State for test by @FlorianDeconinck in #340
• [Feat | Schedule Tree Opt] Axis Merge: push non-cartesian ForScope down by @FlorianDeconinck in #342
• [Feature | Fix] Boilerplate adjustements for physics and bench by @FlorianDeconinck in #343
• [Feat] Better State and LocaL for real model use case by @FlorianDeconinck in #344
• [Feature] Introducing Xumpy by @FlorianDeconinck in #345
• Minor fix to column min/max functions by @CharlesKrop in #349
• BREAKING CHANGE: remove deprecated CopyCornersXY by @romanc in #351
• BREAKING CHANGE: require comm object in ZarrMonitor by @romanc in #352
• refactor: update GitHub workflow dependencies by @romanc in #354
• [Feature | Orchestration] Optimizated C call overhead away, update pipeline, optimize CPU transient residency by @FlorianDeconinck in #348
• [Update] GT4Py up to January 7th develop by @FlorianDeconinck in #355
• Parameterize SubtileGridSizer to allow variation of the padding on non-interface field by @FlorianDeconinck in #347
• refactor: deprecate duplicate basic stencil definitions by @romanc in #358
• [Feature | Experimental] SDFG code labeler by @FlorianDeconinck in #359
• BREAKING CHANGE: remove get_tile_number() in partitioner by @romanc in #353
• refactor: small cleanup in translate tests by @romanc in #360
• CI: Add h5py dependency for h5netcdf, which now supports multiple backends by @romanc in #363
• BREAKING CHANGE: remove deprecated QuanityFactory.from_backend() by @romanc in #356
• BREAKING CHANGE: Quantity & QuantityMetadata changes by @romanc in #357
• refactor: Types for non-deprecated corner code by @romanc in #361
• BREAKING CHANGE: remove deprecated NullComm by @romanc in #350
• refactor: improve types in python extension by @romanc in #365
• Release: NDSL 2026.01.00 by @romanc in #364

Full Changelog: 2025.11.00...2026.01.00