Milestone M2.5: IFS-FVM can run with GT4Py on unstructured grids on the sphere

This milestone demonstrates that PMAP, the Python implementation of the IFS finite-volume dynamical core (IFS-FVM) built on the GT4Py domain-specific language (DSL), can run on unstructured grids on the sphere. The implementation uses GT4Py.next, the latest version of GT4Py, which supports unstructured grids and is still under active development. The global PMAP configuration uses the horizontally unstructured, vertically structured octahedral grid of the operational IFS and has been validated against the original Fortran IFS-FVM using the DCMIP-2016 baroclinic instability benchmark, showing close agreement in the simulated atmospheric flow after a 15-day forecast (ESiWACE3 D2.3). GT4Py.next generates optimised CPU and GPU code from the same high-level Python operators, so that the global and regional PMAP configurations share one programming model. Weak-scaling experiments on the Alps system at CSCS, using NVIDIA Grace Hopper GH200 GPUs, have also demonstrated good scalability (ESiWACE3 D2.4). These results confirm the suitability of GT4Py for global atmospheric modelling and its potential to support portable, high-performance implementations of numerical weather prediction models on modern HPC architectures.The global PMAP configuration has also been used to investigate numerical precision reduction and mixed-precision configurations (ESiWACE3 D2.2).GT4Py is released under the BSD-3-Clause licence, and the precision-reduction extension is publicly available through the GitHub repository https://github.com/GridTools/gt4py. A public release of PMAP under the Apache-2.0 licence is planned at ECMWF for 2027.