AIFS ENS v2: ECMWF's data-driven ensemble weather forecasting model

ECMWF's Artificial Intelligence Forecasting System (AIFS) ENS v2,
implemented operationally on 12 May 2026, superseding v1. Generates
51-member 15-day (6-hourly) global ensemble forecasts four times per day.

Built on a GNN encoder/decoder with sliding window transformer processor,
trained on ERA5 (1979-2022) and fine-tuned on operational analysis data
(2018-2024). Trained on 32 GH200 GPUs using the Anemoi framework.

Key features in v2:
- First operational data-driven wave forecasts (11 wave variables)
- New snow variable (fraction of snow cover)
- Tropical cyclone track forecasts (BUFR format)
- Improved stratosphere representation (10hPa pressure level)
- Improved vertical velocities (W changed to diagnostic field)
- Multi-scale loss replacing afCRPS loss function
- Revised graph features with more decoder edges
- Variable bounding matching AIFS Single for physical consistency

Includes pretrained checkpoint (2.55GB), inference notebook, and
uv lock files for reproducible environments.

Co-authored-by: Meghan Plumridge <maplumridge@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,39 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.grib filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/aifs_diagram.png filter=lfs diff=lfs merge=lfs -text
+assets/decoder_graph.jpeg filter=lfs diff=lfs merge=lfs -text
+assets/encoder_graph.jpeg filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1 @@
+.venv

.pre-commit-config.yaml

@@ -0,0 +1,48 @@
+repos:
+- repo: https://github.com/pre-commit/pre-commit-hooks
+  rev: v6.0.0
+  hooks:
+  - id: check-yaml # Check YAML files for syntax errors only
+    args: [--unsafe, --allow-multiple-documents]
+  - id: debug-statements # Check for debugger imports and py37+ breakpoint()
+  - id: end-of-file-fixer # Ensure files end in a newline
+  - id: trailing-whitespace # Trailing whitespace checker
+  - id: check-merge-conflict # Check for files that contain merge conflict
+- repo: https://github.com/pre-commit/pygrep-hooks
+  rev: v1.10.0  # Use the ref you want to point at
+  hooks:
+  - id: python-use-type-annotations # Check for missing type annotations
+  - id: python-check-blanket-noqa # Check for # noqa: all
+  - id: python-no-log-warn # Check for log.warn
+- repo: https://github.com/psf/black-pre-commit-mirror
+  rev: 26.1.0
+  hooks:
+  - id: black
+    args: [--line-length=120]
+- repo: https://github.com/pycqa/isort
+  rev: 8.0.1
+  hooks:
+  - id: isort
+    args:
+    - -l 120
+    - --force-single-line-imports
+    - --profile black
+    - --project anemoi
+- repo: https://github.com/astral-sh/ruff-pre-commit
+  rev: v0.15.4
+  hooks:
+  - id: ruff
+    args:
+    - --line-length=120
+    - --fix
+    - --exit-non-zero-on-fix
+    - --exclude=docs/**/*_.py
+- repo: https://github.com/sphinx-contrib/sphinx-lint
+  rev: v1.0.2
+  hooks:
+  - id: sphinx-lint
+- repo: https://github.com/tox-dev/pyproject-fmt
+  rev: "v2.16.2"
+  hooks:
+  - id: pyproject-fmt
+    args: ["--max-supported-python", "3.12"]

README.md

@@ -0,0 +1,304 @@
+---
+license: cc-by-4.0
+pipeline_tag: graph-ml
+language:
+- en
+library_name: anemoi
+---
+
+# AIFS ENS v2
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+[![License](https://img.shields.io/badge/license-CC--BY--4.0-blue.svg)](https://creativecommons.org/licenses/by/4.0/)
+[![Framework](https://img.shields.io/badge/framework-Anemoi-orange)](https://anemoi.readthedocs.io/)
+[![Task](https://img.shields.io/badge/task-Weather%20Forecasting-green)]()
+[![Model](https://img.shields.io/badge/model-GNN%20%2B%20Transformer-purple)]()
+
+The AIFS is ECMWF's **Artificial Intelligence Forecasting System**.
+It consists of two data-driven medium-range forecast models:
+- AIFS Single v2 (described on the [following page](https://huggingface.co/ecmwf/aifs-single-2.0))
+- AIFS ENS v2 (described here)
+
+**AIFS ENS v2** was implemented on **12 May 2026** and supersedes version [1](https://huggingface.co/ecmwf/aifs-ens-1).
+It is run operationally by ECMWF, generating a 51-member 15-day (6-hourly) global forecast four times per day.
+
+## Table of contents
+
+- [What's new in v2](#whats-new-in-v2)
+- [Quickstart](#quickstart)
+- [Model overview](#model-overview)
+- [Data details](#data-details)
+- [Training details](#training-details)
+- [Evaluation](#evaluation)
+- [Known limitations](#known-limitations)
+- [Technical specifications](#technical-specifications)
+- [Citation](#citation)
+
+---
+
+## What's new in v2
+The release of AIFS v2 introduces:
+- A new wave component, including 11 wave variables, marking **ECMWF's first operational data-driven wave forecasts**.
+- The addition of a new snow variable to the existing land component.
+- The addition of tropical cyclone track forecasts.
+- The addition of two variables already present in AIFS Single (vsw and cp), to harmonise the two models.
+- Improved vertical velocities, by changing parameter W from a prognostic to a diagnostic field.
+- An improved representation of the stratosphere, with the addition of pressure level fields at 10hPa.
+
+Architectural changes include:
+- Replacing the afCRPS loss function used in AIFS ENS [v1](https://huggingface.co/ecmwf/aifs-ens-1) with a multi-scale loss.
+- Imposing the same variable bounding as is used by the AIFS Single model, improving physical consistency.
+- Revised graph features, utilising more edges in the decoder.
+- An improved training regime, using 1 year of additional training data compared to AIFS ENS [v1](https://huggingface.co/ecmwf/aifs-ens-1).
+
+For full details of the changes introduced with this upgrade, see the [implementation page](https://confluence.ecmwf.int/display/FCST/Implementation+of+AIFS+ENS+v2).
+
+---
+
+## Quickstart
+### Access AIFS Single v2 model data
+ECMWF generates operational forecasts from AIFS ENS v2 four times per day (at 00, 06, 12 & 18 UTC).
+Users can access the forecast data free-of-charge through various [open data platforms](https://www.ecmwf.int/en/forecasts/datasets/open-data).
+
+### Generate a forecast with AIFS Single v2
+To generate a forecast using the AIFS ENS v2 model, follow the [notebook example](run_AIFS_ENS_v2.0.ipynb).
+
+The notebook demonstrates:
+- installing packages and imports
+- retrieving initial conditions from ECMWF Open Data
+- loading the pretrained checkpoint
+- running inference with [anemoi-inference](https://github.com/ecmwf/anemoi-inference)
+- visualising the forecast
+
+[anemoi-inference](https://github.com/ecmwf/anemoi-inference) also provides a command line interface:
+```bash
+anemoi-inference run inference.yaml
+```
+
+or, if using uv from this repository:
+```bash
+uv run --extra inference anemoi-inference run inference.yaml
+```
+
+---
+
+## Model overview
+
+### Model description
+
+<!-- Provide a longer summary of what this model is. -->
+
+AIFS ENS v2 is based on a graph neural network (GNN) encoder and decoder, and a sliding window transformer processor.
+
+<div style="display: flex; justify-content: center;">
+  <img src="assets/encoder_graph.jpeg" alt="Encoder graph" style="width: 50%;"/>
+  <img src="assets/decoder_graph.jpeg" alt="Decoder graph" style="width: 50%;"/>
+</div>
+
+The model has a flexible and modular design and supports several levels of parallelism to enable training on
+high resolution input data. AIFS forecast skill is assessed by comparing its forecasts to numerical weather prediciton (NWP) analyses
+and direct observational data.
+
+- **Developed by:** ECMWF
+- **Model type:** Encoder-processor-decoder model
+- **License:** Inference model weights are published under a Creative Commons Attribution 4.0 International (CC BY 4.0).
+To view a copy of this licence, visit https://creativecommons.org/licenses/by/4.0/.
+The [demonstration notebook](run_AIFS_ENS_v2.0.ipynb) and other script files are published under an Apache 2.0 licence.
+To view a copy of this license, visit https://www.apache.org/licenses/LICENSE-2.0.txt.
+
+### Model resolution
+AIFS ENS v2 introduces a **new pressure level** to the stratospheric component.
+| Model | Vertical resolution [pressure levels] (hPa) |
+|:---|:---|
+| AIFS ENS v2 | 10 (new), 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000 |
+
+There are no changes in horiztonal resolution compared to previous version AIFS ENS v2.
+| Component | Horizontal Resolution [kms] | Vertical Resolution [levels] |
+|---|:---:|:---:|
+| Atmosphere | ~ 31 |  13 |
+
+### Model sources
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [Anemoi](https://anemoi.readthedocs.io/en/latest/) is an open-source framework for
+  creating data-driven weather forecasting systems. Anemoi is co-developed by ECMWF and national meteorological
+  services across Europe.
+- **Papers:**
+  - [AIFS-CRPS: Ensemble forecasting using a model trained with a loss function based on the Continuous Ranked Probability Score (2024)](https://arxiv.org/abs/2412.15832)
+  - [A multi-scale loss formulation for learning a probabilistic model with proper score optimisation (2025)](https://arxiv.org/abs/2506.10868)
+
+---
+
+## Data details
+
+### Training data
+The data used for **pre-training** AIFS ENS v2 remains the same as for the previous model version ([v1](https://huggingface.co/ecmwf/aifs-ens-1)):
+- Pre-training was performed on ERA5 data covering the years 1979–2022 over 300,000 training steps.
+
+The data used for **fine-tuning** AIFS Single v2 has been updated:
+- Fine-tuning was performed on 1 year of additional data compared to [AIFS ENS v1](https://huggingface.co/ecmwf/aifs-ens-1).
+- In particular, fine-tuning was performed on ECMWF operational analysis data and
+  IFS 50r1 esuite analysis data, covering the years 2018-2024 over 7,900 training steps.
+
+<div style="display: flex; justify-content: center;">
+  <img src="assets/ens_v2_training.png" alt="AIFS ENS v2 fine-tuning" style="width: 80%;"/>
+</div>
+
+> *Note: The IFS 50r1 esuite analysis data used for fine-tuning is not available to users. It consists of prototype data from early versions of IFS Cycle 50r1.*
+
+As in previous versions of AIFS ENS, IFS fields are interpolated from their native O1280 resolution
+(approximately 0.1°) using MARS default interpolation tools down to N320 (approximately 0.25°) for
+fine-tuning and initialisation of the model during inference.
+
+### Data parameters
+
+#### New parameters
+
+AIFS Single v2 introduces 14 new parameters, which were used for model training and are output by the
+model during a forecast.
+
+| Short Name | Name | Units |
+|:----------:|:----:|:-----:|
+| h1012 | Significant wave height of all waves with periods within the inclusive range from 10 to 12 seconds | \(m\) |
+| h1214 | Significant wave height of all waves with periods within the inclusive range from 12 to 14 seconds | \(m\) |
+| h1417 | Significant wave height of all waves with periods within the inclusive range from 14 to 17 seconds | \(m\) |
+| h1721 | Significant wave height of all waves with periods within the inclusive range from 17 to 21 seconds | \(m\) |
+| h2125 | Significant wave height of all waves with periods within the inclusive range from 21 to 25 seconds | \(m\) |
+| h2530 | Significant wave height of all waves with periods within the inclusive range from 25 to 30 seconds | \(m\) |
+| wmb | Model bathymetry | \(m\) |
+| swh | Significant wave height | \(m\) |
+| mwd | Mean wave direction | \(Degree\ true\) |
+| mwp | Mean wave period | \(s\) |
+| cdww | Coefficient of drag with waves | \(dimensionless\) |
+| fscov | Fraction of snow cover | \(Proportion\) |
+| cp | Convective precipitation | \(kg m-2\) |
+| vsw | Volumetric soil moisture | \(m3 m-3\) |
+
+Additionaly, AIFS ENS v2.0 now produces tropical cyclone track forecasts in BUFR format.
+See https://confluence.ecmwf.int/display/FCST/Implementation+of+AIFS+ENS+v2#ImplementationofAIFSENSv2-Newparameters for more details.
+
+---
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+AIFS ENS v2 is trained on ECMWF’s ERA5 re-analysis and ECMWF’s operational numerical weather prediction (NWP) analyses as
+described in the [training data](#training-data) section.
+
+AIFS ENS v2 is trained to produce 6-hour forecasts. It receives as input a representation of the atmospheric states
+at \\(t_{−6h}\\), \\(t_{0}\\), and then forecasts the state at time \\(t_{+6h}\\).
+
+<div style="display: flex; justify-content: center;">
+  <img src="assets/aifs_diagram.png" alt="AIFS 2m Temperature" style="width: 80%;"/>
+</div>
+
+The table below shows all parameters used and output by AIFS ENS v2. New parameters and levels are marked **bold**.
+
+| Field | Level type | Input/Output |
+|---|---|---|
+| Geopotential (Z), horizontal and vertical wind components (U, V), temperature (T) | Pressure levels: **10**, 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000 | Both ("Prognostic") |
+| Specific humidity (Q) | Pressure levels: 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000 | Both ("Prognostic") |
+| Vertical velocity (W) | Pressure levels: **10**, 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000 | Output ("Diagnostic") |
+| Specific humidity (Q) | Pressure level: 50 | Output ("Diagnostic") |
+| Surface pressure (SP), mean sea-level pressure (MSL), sea-surface temperature (SST), skin temperature (SKT), 2m temperature (2T), 2m dewpoint temperature (2D), 10m horizontal wind components (10U, 10V), total column water (TCW), **mean wave period (MWP)**, **mean wave direction (MWD)**, **coefficient of drag with waves (CDWW)**, **significant wave height (SWH)**, significant wave height of all waves with periods within the inclusive range from: <br><br> - **10 to 12 seconds (H1012)** <br> - **12 to 14 seconds (H1214)** <br> - **14 to 17 seconds (H1417)** <br> - **17 to 21 seconds (H1721)** <br> - **21 to 25 seconds (H2125)** <br> - **25 to 30 seconds (H2530)** | Surface | Both ("Prognostic") |
+| **Volumetric soil moisture (VSW)** and soil temperature (SOT), both at soil depth 1 and 2 | Soil layer | Both ("Prognostic") |
+| 100m horizontal wind components (100U, 100V), surface short-wave (solar) radiation downwards (SSRD), surface long-wave (thermal) radiation downwards (STRD), cloud variables (TCC, HCC, MCC, LCC), runoff water equivalent (ROWE) and snow fall (SF), total precipitation (TP), **convective precipitation (CP)**, **fraction of snow cover (FSCOV)** | Surface | Output ("Diagnostic") |
+| Standard deviation of sub-gridscale orography (SDOR), slope of sub-gridscale orography (SLOR), land-sea mask (LSM), Geopotential (Z), insolation, latitude/longitude, time of day/day of year | Surface | Input ("Forcings") |
+
+### Training Procedure
+
+Key changes to the AIFS ENS v2 training regime include:
+- the introduction of a multi-scale loss
+- the removal of reference field truncation, imposing the same variable bounds used by the AIFS Single model to improve physical consistency
+- the revision of graph features, by using more edges in the decoder together with new edge features.
+
+Full details about the model architecture are detailed in the arXiv
+preprints [here](https://arxiv.org/abs/2412.15832) and [here](https://arxiv.org/abs/2506.10868).
+
+#### Speeds, Sizes, Times
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+Data parallelism is used for training, with a batch size of 1. One model instance is split across two 120GB GH200
+GPUs, with an ensemble instance split across 4 gpus. Training is done using mixed precision (Micikevicius et al. [2018]), and the entire process
+takes about one week, with 32 GPUs in total. The checkpoint size is 2.55GB and as mentioned above, it does not include the optimizer
+state.
+
+---
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+Interactive scorecards presenting the performance of AIFS ENS v2 between January-March 2026 are now available.
+The scorecards compare performance when initialised from 49r1 and 50r1 IFS initial conditions: 
+
+- [AIFS ENS v2 (initialised from 50r1) compared with AIFS ENS v1 (initialised from 49r1)](https://sites.ecmwf.int/aifs/scorecards/AIFS-ENS-v2%20vs%20AIFS-ENS-v1.html)
+- [AIFS ENS v2 compared with AIFS ENS v1 (both initialised from 50r1)](https://sites.ecmwf.int/aifs/scorecards/AIFS-ENS-v2%20vs%20AIFS-ENS-v1%20from%2050r1.html)
+
+---
+
+## Known limitations
+
+Please refer to https://confluence.ecmwf.int/display/FCST/Known+AIFS+Forecasting+Issues.
+
+---
+
+## Technical specifications
+
+### Hardware
+
+<!--  {{ hardware_requirements | default("[More Information Needed]", true)}} -->
+
+AIFS ENS v2 was trained on 32 GH200 GPUs (120GB).
+
+### Software
+
+The model was developed and trained using the [Anemoi framework](https://anemoi.readthedocs.io/en/latest/).
+The Anemoi framework provides a complete toolkit to develop data-driven weather models –
+from data preparation through to inference. The development is primarily driven by a number
+of European Meterological Organisations but open to contributions from any organisation or any individual.
+The framework is composed of several packages which target the different components necessary to
+construct data-driven weather models. To aid development and deployment, each package collects metadata
+that can be used by the subsequent packages. The framework builds upon on established Python tools
+including PyTorch, Lighting, Hydra, Zarr, Xarray and earthkit.
+
+---
+
+## Citation
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+If you use this model in your work, please cite it as follows:
+
+**BibTeX:**
+
+```bibtex
+@misc{lang2025multiscale,
+      title={A multi-scale loss formulation for learning a probabilistic model with proper score optimisation},
+      author={Simon Lang and Martin Leutbecher and Pedro Maciel},
+      year={2025},
+      eprint={2506.10868},
+      archivePrefix={arXiv},
+      primaryClass={physics.ao-ph},
+      url={https://arxiv.org/abs/2506.10868},
+}
+
+```
+
+**APA:**
+
+```apa
+Lang, S., Leutbecher, M., & Maciel, P. (2025). A multi-scale loss formulation for learning a probabilistic model with proper score optimisation. arXiv preprint arXiv:2506.10868.
+
+```
+
+## More Information
+
+All papers:
+- [AIFS-CRPS: Ensemble forecasting using a model trained with a loss function based on the Continuous Ranked Probability Score (2024)](https://arxiv.org/abs/2412.15832)
+- [A multi-scale loss formulation for learning a probabilistic model with proper score optimisation (2025)](https://arxiv.org/abs/2506.10868)

aifs-ens-crps-2.0.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e383027ee9a20469b920ed11f9af2ecc9268ebcf826447638c5c9fa296fbbbdb
+size 2553843230

inference.yaml

@@ -0,0 +1,56 @@
+description: |
+  AIFS-ens CRPS 2.0
+
+checkpoint: aifs-ens-crps-2.0.ckpt
+
+input: opendata
+
+pre_processors:
+  - forward-transform-filter: cos_sin_mean_wave_direction
+  - forward-transform-filter: # Mask fix
+      filter: apply-mask
+      param:
+        - sd
+        - swvl1
+        - swvl2
+      path: lsm.grib
+      mask_value: 0
+
+post_processors:
+  - backward-transform-filter: cos_sin_mean_wave_direction
+  - accumulate_from_start_of_forecast
+
+patch_metadata:
+  dataset:
+    constant_fields: [z, sdor, slor, lsm]
+
+use_grib_paramid: true
+allow_nans: true
+typed_variables:
+  mwd:
+    mars:
+      param: mwd
+      stream: waef
+      levtype: sfc
+  snowc: # Encoding fix
+    mars:
+      param: fscov
+      stream: enfo
+      levtype: sfc
+  # fscov: # required for step0 copy
+  #   mars:
+  #     param: fscov
+  #     stream: enfo
+  #     levtype: sfc
+
+output:
+  grib:
+    path: output.grib
+    encoding:
+      class: ai
+      type: pf
+      model: aifs-ens
+      generatingProcessIdentifier: 2
+
+env:
+  ANEMOI_INFERENCE_NUM_CHUNKS: ${oc.env:ANEMOI_INFERENCE_NUM_CHUNKS,8}

lsm.grib

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6660ca07729fd8f2ac3fb8bfeee66e7513a355ca90d287f39b5987832c44a891
+size 1627680

pyproject.toml

@@ -0,0 +1,42 @@
+[project]
+name = "aifs-ens"
+version = "2.0"
+requires-python = ">=3.11,<3.14"
+classifiers = [
+  "Programming Language :: Python :: 3 :: Only",
+  "Programming Language :: Python :: 3.11",
+  "Programming Language :: Python :: 3.12",
+  "Programming Language :: Python :: 3.13",
+]
+dependencies = [
+  "anemoi-graphs==0.8.1",
+  "anemoi-models==0.11.2",
+  "anemoi-transform==0.1.16.post2",
+  "anemoi-utils==0.4.37",
+  "earthkit-data<1",
+  "earthkit-plots<1",
+  "flash-attn==2.7.4",
+  "torch==2.7",
+  "torch-geometric==2.6.1",
+]
+optional-dependencies.inference = [
+  "anemoi-inference[huggingface]==0.8.3",
+  "anemoi-plugins-ecmwf-inference[opendata]==0.2.1",
+  "earthkit-regrid==0.5.1",
+  "ecmwf-opendata==0.3.29",
+]
+optional-dependencies.jupyter = [
+  "ipykernel>=7.2",
+]
+optional-dependencies.training = [
+  "anemoi-training==0.8.1",
+]
+
+[tool.uv]
+sources.flash-attn = [
+  { url = "https://github.com/cathalobrien/get-flash-attn/releases/download/v0.1-alpha/flash_attn-2.8.3+cu12torch2.7cxx11abiFALSE-cp312-cp312-linux_x86_64.whl", marker = "platform_machine == 'x86_64' and sys_platform != 'win32'" },
+  { url = "https://github.com/cathalobrien/get-flash-attn/releases/download/v0.1-alpha/flash_attn-2.8.3+cu12torch2.7cxx11abiFALSE-cp312-cp312-linux_aarch64.whl", marker = "platform_machine == 'aarch64' and sys_platform == 'linux'" },
+]
+environments = [
+  "sys_platform == 'linux' and (platform_machine == 'x86_64' or platform_machine == 'aarch64')",
+]