Repository navigation

Tags

training-0.17.2

chore: Release main (#1376)

🤖 Automated Release PR

This PR was created by `release-please` to prepare the next release.
Once merged:

1. A new version tag will be created
2. A GitHub release will be published
3. The changelog will be updated

Changes to be included in the next release:
---


<details><summary>training: 0.17.2</summary>

##
[0.17.2](training-0.17.0...training-0.17.2)
(2026-11-03)


### ⚠ BREAKING CHANGES

* **training:** integrate the checkpoint pipeline as the single
checkpoint surface
([#495](#495)) (#1196)
* **training:** ANEMOI_BASE_SEED is no longer read; set training.seed.
persistent_workers defaults to false. MultiDataset is now a map-style
dataset. We removed worker_init_func. We add a torchdata dependency.

### Features

* **models:** Grouped ffts and explicitly backward for gaussian grids
([#1414](#1414))
([0bdabce](0bdabce))
* Spatial preprocessor
([#1305](#1305))
([89f26fa](89f26fa))
* Support regular lat-lon grids in spherical harmonic transforms
([#1420](#1420))
([6ad6bd6](6ad6bd6))
* **training:** Incremental mlflow sync
([#1467](#1467))
([e0d2467](e0d2467))
* **training:** Integrate the checkpoint pipeline as the single
checkpoint surface
([#495](#495))
([#1196](#1196))
([aa4d3f1](aa4d3f1))
* **training:** Keep track of changes in the accuracy test and make it
easier to run with arbitrary references
([#1463](#1463))
([33fe303](33fe303))
* **training:** Mlflow sync uses the run's max_params_length
([#1464](#1464))
([b6295c5](b6295c5))
* **training:** Resume training exactly, including mid-epoch, map style
dataset ([#1433](#1433))
([d252ca4](d252ca4))


### Bug Fixes

* Add task to metadata inference
([#1450](#1450))
([9ced7fa](9ced7fa))
* **graphs:** Pyg-lib support
([#1407](#1407))
([6ae7780](6ae7780))
* **tests:** Make test config loading independent of the compiled file
name ([#1466](#1466))
([def4553](def4553))
* **tests:** Make tests discoverable
([#1359](#1359))
([54db8ce](54db8ce))
* **training, graphs:** Small fixes
([#1424](#1424))
([6a6f042](6a6f042))
* **training,models,graphs:** Bump torch version 2.11
([#1455](#1455))
([e21aefd](e21aefd))
* **training:** Anemoi-training checkpoint inference
([#1490](#1490))
([00a5a27](00a5a27))
* **training:** Benchmark tests run again
([#1382](#1382))
([3594d55](3594d55))
* **training:** Check for initialise_data_extractor_zero
([#1447](#1447))
([91bdd53](91bdd53))
* **training:** Fix compile error when using activation checkpointing
([#1344](#1344))
([10f8c25](10f8c25))
* **training:** Fix edge trainable parameteres plotting callback
([#1482](#1482))
([ec560b8](ec560b8))
* **training:** Import datashader lazily
([#1448](#1448))
([07513e3](07513e3))
* **training:** Match validation loss to training rollout.
([#1384](#1384))
([997d6a6](997d6a6))
* **training:** Replace memory efficient but unstable graph edge energ
score formulation
([#1354](#1354))
([d7519f6](d7519f6))
* **training:** Resolve subgrid
([#1363](#1363))
([5fae1b9](5fae1b9))
* **training:** Scale prog variable at data index 0 in the tendency
scaler ([#1369](#1369))
([a5d9693](a5d9693))
* **training:** Scalers and target variable
([#1371](#1371))
([457a8c6](457a8c6))
* **training:** Sharded ensemble gradient scaling
([#1377](#1377))
([23c8bfd](23c8bfd))
* **training:** Update transformer.yaml
([#1389](#1389))
([396e1bb](396e1bb))


### Performance Improvements

* Models, training:trim checkpoint
([#1456](#1456))
([42260df](42260df))
* **models,training:** Triton-transformer-backend
([#1444](#1444))
([6636d69](6636d69))
* **training:** Fake dataloading
([#1387](#1387))
([53dcf12](53dcf12))
* **training:** Improving loss and data reader performance
([#1469](#1469))
([3452dec](3452dec))


### Documentation

* **training:** Correct anemoi-datasets ERA5 Zarr copy cmd
([#1489](#1489))
([629dec8](629dec8))
* **training:** Expand Data-Distributed section with multi-GPU usage
([#1390](#1390))
([b963c36](b963c36))
* **training:** Fix stale config paths in user-guide
([#1391](#1391))
([42c8a06](42c8a06))
* **training:** Link User Journey items in user-guide overview
([#1388](#1388))
([08b0a64](08b0a64))
* **training:** Replace deprecated AlmostFairKernelCRPS references wi…
([#1392](#1392))
([f90ec38](f90ec38))


### Miscellaneous Chores

* Bump training 0.17.2
([#1438](#1438))
([173cd2d](173cd2d))
</details>

<details><summary>graphs: 0.9.9</summary>

##
[0.9.9](graphs-0.9.7...graphs-0.9.9)
(2026-11-03)


### Features

* **graphs:** HEALPix Node Ordering Schema Option and kwarg
([#1381](#1381))
([a4ba37d](a4ba37d))


### Bug Fixes

* **graphs:** Have inspect load the graph on cpu
([#1487](#1487))
([4090ee8](4090ee8))
* **graphs:** Pin xarray
([#1436](#1436))
([47c2f3e](47c2f3e))
* **graphs:** Pyg-lib support
([#1407](#1407))
([6ae7780](6ae7780))
* **tests:** Make tests discoverable
([#1359](#1359))
([54db8ce](54db8ce))
* Torch2.6 compatibility
([#1383](#1383))
([a48ce21](a48ce21))
* **training, graphs:** Small fixes
([#1424](#1424))
([6a6f042](6a6f042))
* **training,models,graphs:** Bump torch version 2.11
([#1455](#1455))
([e21aefd](e21aefd))


### Documentation

* **graphs:** Fix cut-off equations and installing include
([#1486](#1486))
([67f1dec](67f1dec))


### Miscellaneous Chores

* Bump graphs 0.9.9
([#1439](#1439))
([772c798](772c798))
</details>

<details><summary>models: 0.20.0</summary>

##
[0.20.0](models-0.19.0...models-0.20.0)
(2026-11-03)


### ⚠ BREAKING CHANGES

* **training:** integrate the checkpoint pipeline as the single
checkpoint surface
([#495](#495)) (#1196)
* **models:** remove unused layer norms from TransformerMapperBlock
([#1443](#1443))

### Features

* **models:** Add cross-attention aggregator for dataset latents
([#1309](#1309))
([f390eef](f390eef))
* **models:** Add tests for halo exchange primitives and autograd-aware
communication functions
([#1442](#1442))
([3ab4a11](3ab4a11))
* **models:** Grouped ffts and explicitly backward for gaussian grids
([#1414](#1414))
([0bdabce](0bdabce))
* **models:** Tests for the autograd-aware communication functions in
``anemoi.models.distributed.graph``
([#1299](#1299))
([868fea5](868fea5))
* Spatial preprocessor
([#1305](#1305))
([89f26fa](89f26fa))
* Support regular lat-lon grids in spherical harmonic transforms
([#1420](#1420))
([6ad6bd6](6ad6bd6))
* **training:** Integrate the checkpoint pipeline as the single
checkpoint surface
([#495](#495))
([#1196](#1196))
([aa4d3f1](aa4d3f1))


### Bug Fixes

* Add error when passing multiple datasets for one encoder
([#1394](#1394))
([a01e517](a01e517))
* Docstrings SkipConnection
([#1431](#1431))
([02ca484](02
8000
ca484))
* **models:** Decoder-only multi-dataset training for ensemble models
([#1401](#1401))
([6a98583](6a98583))
* **models:** Extend halo exchange to non-symmetric case
([#1418](#1418))
([2037e6f](2037e6f))
* **models:** Refuse dropout in processors and mappers when the model is
sharded ([#1497](#1497))
([666af83](666af83))
* **models:** Remove unused layer norms from TransformerMapperBlock
([#1443](#1443))
([8c3ab41](8c3ab41))
* **models:** Replace the Triton import check in gt.py
([#1409](#1409))
([fe77248](fe77248))
* **models:** Skip input-only datasets in predict_step post-processing
([#1453](#1453))
([dd9c266](dd9c266))
* **models:** Spatial_pre_processors attribute access
([#1410](#1410))
([b52b034](b52b034))
* **models:** Subset forcings using regular indexing instead of
torch.index_select to fix device mismatch
([#1429](#1429))
([d09c61d](d09c61d))
* **tests:** Make tests discoverable
([#1359](#1359))
([54db8ce](54db8ce))
* Torch2.6 compatibility
([#1383](#1383))
([a48ce21](a48ce21))
* **training,models,graphs:** Bump torch version 2.11
([#1455](#1455))
([e21aefd](e21aefd))


### Performance Improvements

* **models,training:** Triton-transformer-backend
([#1444](#1444))
([6636d69](6636d69))
</details>

---
> [!IMPORTANT]
> Please do not change the PR title, manifest file, or any other
automatically generated content in this PR unless you understand the
implications. Changes here can break the release process.
> 
> ⚠️ Merging this PR will:
> - Create a new release
> - Trigger deployment pipelines
> - Update package versions

 **Before merging:**
 - Ensure all tests pass
 - Review the changelog carefully
 - Get required approvals

[Release-please
documentation](https://github.com/googleapis/release-please)

Nov 3, 2026
6a2377d
zip
tar.gz
Notes

models-0.20.0

chore: Release main (#1376)

🤖 Automated Release PR

This PR was created by `release-please` to prepare the next release.
Once merged:

1. A new version tag will be created
2. A GitHub release will be published
3. The changelog will be updated

Changes to be included in the next release:
---


<details><summary>training: 0.17.2</summary>

##
[0.17.2](training-0.17.0...training-0.17.2)
(2026-11-03)


### ⚠ BREAKING CHANGES

* **training:** integrate the checkpoint pipeline as the single
checkpoint surface
([#495](#495)) (#1196)
* **training:** ANEMOI_BASE_SEED is no longer read; set training.seed.
persistent_workers defaults to false. MultiDataset is now a map-style
dataset. We removed worker_init_func. We add a torchdata dependency.

### Features

* **models:** Grouped ffts and explicitly backward for gaussian grids
([#1414](#1414))
([0bdabce](0bdabce))
* Spatial preprocessor
([#1305](#1305))
([89f26fa](89f26fa))
* Support regular lat-lon grids in spherical harmonic transforms
([#1420](#1420))
([6ad6bd6](6ad6bd6))
* **training:** Incremental mlflow sync
([#1467](#1467))
([e0d2467](e0d2467))
* **training:** Integrate the checkpoint pipeline as the single
checkpoint surface
([#495](#495))
([#1196](#1196))
([aa4d3f1](aa4d3f1))
* **training:** Keep track of changes in the accuracy test and make it
easier to run with arbitrary references
([#1463](#1463))
([33fe303](33fe303))
* **training:** Mlflow sync uses the run's max_params_length
([#1464](#1464))
([b6295c5](b6295c5))
* **training:** Resume training exactly, including mid-epoch, map style
dataset ([#1433](#1433))
([d252ca4](d252ca4))


### Bug Fixes

* Add task to metadata inference
([#1450](#1450))
([9ced7fa](9ced7fa))
* **graphs:** Pyg-lib support
([#1407](#1407))
([6ae7780](6ae7780))
* **tests:** Make test config loading independent of the compiled file
name ([#1466](#1466))
([def4553](def4553))
* **tests:** Make tests discoverable
([#1359](#1359))
([54db8ce](54db8ce))
* **training, graphs:** Small fixes
([#1424](#1424))
([6a6f042](6a6f042))
* **training,models,graphs:** Bump torch version 2.11
([#1455](#1455))
([e21aefd](e21aefd))
* **training:** Anemoi-training checkpoint inference
([#1490](#1490))
([00a5a27](00a5a27))
* **training:** Benchmark tests run again
([#1382](#1382))
([3594d55](3594d55))
* **training:** Check for initialise_data_extractor_zero
([#1447](#1447))
([91bdd53](91bdd53))
* **training:** Fix compile error when using activation checkpointing
([#1344](#1344))
([10f8c25](10f8c25))
* **training:** Fix edge trainable parameteres plotting callback
([#1482](#1482))
([ec560b8](ec560b8))
* **training:** Import datashader lazily
([#1448](#1448))
([07513e3](07513e3))
* **training:** Match validation loss to training rollout.
([#1384](#1384))
([997d6a6](997d6a6))
* **training:** Replace memory efficient but unstable graph edge energ
score formulation
([#1354](#1354))
([d7519f6](d7519f6))
* **training:** Resolve subgrid
([#1363](#1363))
([5fae1b9](5fae1b9))
* **training:** Scale prog variable at data index 0 in the tendency
scaler ([#1369](#1369))
([a5d9693](a5d9693))
* **training:** Scalers and target variable
([#1371](#1371))
([457a8c6](457a8c6))
* **training:** Sharded ensemble gradient scaling
([#1377](#1377))
([23c8bfd](23c8bfd))
* **training:** Update transformer.yaml
([#1389](#1389))
([396e1bb](396e1bb))


### Performance Improvements

* Models, training:trim checkpoint
([#1456](#1456))
([42260df](42260df))
* **models,training:** Triton-transformer-backend
([#1444](#1444))
([6636d69](6636d69))
* **training:** Fake dataloading
([#1387](#1387))
([53dcf12](53dcf12))
* **training:** Improving loss and data reader performance
([#1469](#1469))
([3452dec](3452dec))


### Documentation

* **training:** Correct anemoi-datasets ERA5 Zarr copy cmd
([#1489](#1489))
([629dec8](629dec8))
* **training:** Expand Data-Distributed section with multi-GPU usage
([#1390](#1390))
([b963c36](b963c36))
* **training:** Fix stale config paths in user-guide
([#1391](#1391))
([42c8a06](42c8a06))
* **training:** Link User Journey items in user-guide overview
([#1388](#1388))
([08b0a64](08b0a64))
* **training:** Replace deprecated AlmostFairKernelCRPS references wi…
([#1392](#1392))
([f90ec38](f90ec38))


### Miscellaneous Chores

* Bump training 0.17.2
([#1438](#1438))
([173cd2d](173cd2d))
</details>

<details><summary>graphs: 0.9.9</summary>

##
[0.9.9](graphs-0.9.7...graphs-0.9.9)
(2026-11-03)


### Features

* **graphs:** HEALPix Node Ordering Schema Option and kwarg
([#1381](#1381))
([a4ba37d](a4ba37d))


### Bug Fixes

* **graphs:** Have inspect load the graph on cpu
([#1487](#1487))
([4090ee8](4090ee8))
* **graphs:** Pin xarray
([#1436](#1436))
([47c2f3e](47c2f3e))
* **graphs:** Pyg-lib support
([#1407](#1407))
([6ae7780](6ae7780))
* **tests:** Make tests discoverable
([#1359](#1359))
([54db8ce](54db8ce))
* Torch2.6 compatibility
([#1383](#1383))
([a48ce21](a48ce21))
* **training, graphs:** Small fixes
([#1424](#1424))
([6a6f042](6a6f042))
* **training,models,graphs:** Bump torch version 2.11
([#1455](#1455))
([e21aefd](e21aefd))


### Documentation

* **graphs:** Fix cut-off equations and installing include
([#1486](#1486))
([67f1dec](67f1dec))


### Miscellaneous Chores

* Bump graphs 0.9.9
([#1439](#1439))
([772c798](772c798))
</details>

<details><summary>models: 0.20.0</summary>

##
[0.20.0](models-0.19.0...models-0.20.0)
(2026-11-03)


### ⚠ BREAKING CHANGES

* **training:** integrate the checkpoint pipeline as the single
checkpoint surface
([#495](#495)) (#1196)
* **models:** remove unused layer norms from TransformerMapperBlock
([#1443](#1443))

### Features

* **models:** Add cross-attention aggregator for dataset latents
([#1309](#1309))
([f390eef](f390eef))
* **models:** Add tests for halo exchange primitives and autograd-aware
communication functions
([#1442](#1442))
([3ab4a11](3ab4a11))
* **models:** Grouped ffts and explicitly backward for gaussian grids
([#1414](#1414))
([0bdabce](0bdabce))
* **models:** Tests for the autograd-aware communication functions in
``anemoi.models.distributed.graph``
([#1299](#1299))
([868fea5](868fea5))
* Spatial preprocessor
([#1305](#1305))
([89f26fa](89f26fa))
* Support regular lat-lon grids in spherical harmonic transforms
([#1420](#1420))
([6ad6bd6](6ad6bd6))
* **training:** Integrate the checkpoint pipeline as the single
checkpoint surface
([#495](#495))
([#1196](#1196))
([aa4d3f1](aa4d3f1))


### Bug Fixes

* Add error when passing multiple datasets for one encoder
([#1394](#1394))
([a01e517](a01e517))
* Docstrings SkipConnection
([#1431](#1431))
([02ca484](02ca484))
* **models:** Decoder-only multi-dataset training for ensemble models
([#1401](#1401))
([6a98583](6a98583))
* **models:** Extend halo exchange to non-symmetric case
([#1418](#1418))
([2037e6f](2037e6f))
* **models:** Refuse dropout in processors and mappers when the model is
sharded ([#1497](#1497))
([666af83](666af83))
* **models:** Remove unused layer norms from TransformerMapperBlock
([#1443](#1443))
([8c3ab41](8c3ab41))
* **models:** Replace the Triton import check in gt.py
([#1409](#1409))
([fe77248](fe77248))
* **models:** Skip input-only datasets in predict_step post-processing
([#1453](#1453))
([dd9c266](dd9c266))
* **models:** Spatial_pre_processors attribute access
([#1410](#1410))
([b52b034](b52b034))
* **models:** Subset forcings using regular indexing instead of
torch.index_select to fix device mismatch
([#1429](#1429))
([d09c61d](d09c61d))
* **tests:** Make tests discoverable
([#1359](#1359))
([54db8ce](54db8ce))
* Torch2.6 compatibility
([#1383](#1383))
([a48ce21](a48ce21))
* **training,models,graphs:** Bump torch version 2.11
([#1455](#1455))
([e21aefd](e21aefd))


### Performance Improvements

* **models,training:** Triton-transformer-backend
([#1444](#1444))
([6636d69](6636d69))
</details>

---
> [!IMPORTANT]
> Please do not change the PR title, manifest file, or any other
automatically generated content in this PR unless you understand the
implications. Changes here can break the release process.
> 
> ⚠️ Merging this PR will:
> - Create a new release
> - Trigger deployment pipelines
> - Update package versions

 **Before merging:**
 - Ensure all tests pass
 - Review the changelog carefully
 - Get required approvals

[Release-please
documentation](https://github.com/googleapis/release-please)

Nov 3, 2026
6a2377d
zip
tar.gz
Notes

graphs-0.9.9

chore: Release main (#1376)

🤖 Automated Release PR

This PR was created by `release-please` to prepare the next release.
Once merged:

1. A new version tag will be created
2. A GitHub release will be published
3. The changelog will be updated

Changes to be included in the next release:
---


<details><summary>training: 0.17.2</summary>

##
[0.17.2](training-0.17.0...training-0.17.2)
(2026-11-03)


### ⚠ BREAKING CHANGES

* **training:** integrate the checkpoint pipeline as the single
checkpoint surface
([#495](#495)) (#1196)
* **training:** ANEMOI_BASE_SEED is no longer read; set training.seed.
persistent_workers defaults to false. MultiDataset is now a map-style
dataset. We removed worker_init_func. We add a torchdata dependency.

### Features

* **models:** Grouped ffts and explicitly backward for gaussian grids
([#1414](#1414))
([0bdabce](0bdabce))
* Spatial preprocessor
([#1305](#1305))
([89f26fa](89f26fa))
* Support regular lat-lon grids in spherical harmonic transforms
([#1420](#1420))
([6ad6bd6](6ad6bd6))
* **training:** Incremental mlflow sync
([#1467](#1467))
([e0d2467](e0d2467))
* **training:** Integrate the checkpoint pipeline as the single
checkpoint surface
([#495](#495))
([#1196](#1196))
([aa4d3f1](aa4d3f1))
* **training:** Keep track of changes in the accuracy test and make it
easier to run with arbitrary references
([#1463](#1463))
([33fe303](33fe303))
* **training:** Mlflow sync uses the run's max_params_length
([#1464](#1464))
([b6295c5](b6295c5))
* **training:** Resume training exactly, including mid-epoch, map style
dataset ([#1433](#1433))
([d252ca4](d252ca4))


### Bug Fixes

* Add task to metadata inference
([#1450](#1450))
([9ced7fa](9ced7fa))
* **graphs:** Pyg-lib support
([#1407](#1407))
([6ae7780](6ae7780))
* **tests:** Make test config loading independent of the compiled file
name ([#1466](#1466))
([def4553](def4553))
* **tests:** Make tests discoverable
([#1359](#1359))
([54db8ce](54db8ce))
* **training, graphs:** Small fixes
([#1424](#1424))
([6a6f042](6a6f042))
* **training,models,graphs:** Bump torch version 2.11
([#1455](#1455))
([e21aefd](e21aefd))
* **training:** Anemoi-training checkpoint inference
([#1490](#1490))
([00a5a27](00a5a27))
* **training:** Benchmark tests run again
([#1382](#1382))
([3594d55](3594d55))
* **training:** Check for initialise_data_extractor_zero
([#1447](#1447))
([91bdd53](91bdd53))
* **training:** Fix compile error when using activation checkpointing
([#1344](#1344))
([10f8c25](10f8c25))
* **training:** Fix edge trainable parameteres plotting callback
([#1482](#1482))
([ec560b8](ec560b8))
* **training:** Import datashader lazily
([#1448](#1448))
([07513e3](07513e3))
* **training:** Match validation loss to training rollout.
([#1384](#1384))
([997d6a6](997d6a6))
* **training:** Replace memory efficient but unstable graph edge energ
score formulation
([#1354](#1354))
([d7519f6](d7519f6))
* **training:** Resolve subgrid
([#1363](#1363))
([5fae1b9](5fae1b9))
* **training:** Scale prog variable at data index 0 in the tendency
scaler ([#1369](#1369))
([a5d9693](a5d9693))
* **training:** Scalers and target variable
([#1371](#1371))
([457a8c6](457a8c6))
* **training:** Sharded ensemble gradient scaling
([#1377](#1377))
([23c8bfd](23c8bfd))
* **training:** Update transformer.yaml
([#1389](#1389))
([396e1bb](396e1bb))


### Performance Improvements

* Models, training:trim checkpoint
([#1456](#1456))
([42260df](42260df))
* **models,training:** Triton-transformer-backend
([#1444](#1444))
([6636d69](6636d69))
* **training:** Fake dataloading
([#1387](#1387))
([53dcf12](53dcf12))
* **training:** Improving loss and data reader performance
([#1469](#1469))
([3452dec](3452dec))


### Documentation

* **training:** Correct anemoi-datasets ERA5 Zarr copy cmd
([#1489](#1489))
([629dec8](629dec8))
* **training:** Expand Data-Distributed section with multi-GPU usage
([#1390](#1390))
([b963c36](b963c36))
* **training:** Fix stale config paths in user-guide
([#1391](#1391))
([42c8a06](42c8a06))
* **training:** Link User Journey items in user-guide overview
([#1388](#1388))
([08b0a64](08b0a64))
* **training:** Replace deprecated AlmostFairKernelCRPS references wi…
([#1392](#1392))
([f90ec38](f90ec38))


### Miscellaneous Chores

* Bump training 0.17.2
([#1438](#1438))
([173cd2d](173cd2d))
</details>

<details><summary>graphs: 0.9.9</summary>

##
[0.9.9](graphs-0.9.7...graphs-0.9.9)
(2026-11-03)


### Features

* **graphs:** HEALPix Node Ordering Schema Option and kwarg
([#1381](#1381))
([a4ba37d](a4ba37d))


### Bug Fixes

* **graphs:** Have inspect load the graph on cpu
([#1487](#1487))
([4090ee8](4090ee8))
* **graphs:** Pin xarray
([#1436](#1436))
([47c2f3e](47c2f3e))
* **graphs:** Pyg-lib support
([#1407](#1407))
([6ae7780](6ae7780))
* **tests:** Make tests discoverable
([#1359](#1359))
([54db8ce](54db8ce))
* Torch2.6 compatibility
([#1383](#1383))
([a48ce21](a48ce21))
* **training, graphs:** Small fixes
([#1424](#1424))
([6a6f042](6a6f042))
* **training,models,graphs:** Bump torch version 2.11
([#1455](#1455))
([e21aefd](e21aefd))


### Documentation

* **graphs:** Fix cut-off equations and installing include
([#1486](#1486))
([67f1dec](67f1dec))


### Miscellaneous Chores

* Bump graphs 0.9.9
([#1439](#1439))
([772c798](772c798))
</details>

<details><summary>models: 0.20.0</summary>

##
[0.20.0](models-0.19.0...models-0.20.0)
(2026-11-03)


### ⚠ BREAKING CHANGES

* **training:** integrate the checkpoint pipeline as the single
checkpoint surface
([#495](#495)) (#1196)
* **models:** remove unused layer norms from TransformerMapperBlock
([#1443](#1443))

### Features

* **models:** Add cross-attention aggregator for dataset latents
([#1309](#1309))
([f390eef](f390eef))
* **models:** Add tests for halo exchange primitives and autograd-aware
communication functions
([#1442](#1442))
([3ab4a11](3ab4a11))
* **models:** Grouped ffts and explicitly backward for gaussian grids
([#1414](#1414))
([0bdabce](0bdabce))
* **models:** Tests for the autograd-aware communication functions in
``anemoi.models.distributed.graph``
([#1299](#1299))
([868fea5](868fea5))
* Spatial preprocessor
([#1305](#1305))
([89f26fa](89f26fa))
* Support regular lat-lon grids in spherical harmonic transforms
([#1420](#1420))
([6ad6bd6](6ad6bd6))
* **training:** Integrate the checkpoint pipeline as the single
checkpoint surface
([#495](#495))
([#1196](#1196))
([aa4d3f1](aa4d3f1))


### Bug Fixes

* Add error when passing multiple datasets for one encoder
([#1394](#1394))
([a01e517](a01e517))
* Docstrings SkipConnection
([#1431](#1431))
([02ca484](02ca484))
* **models:** Decoder-only multi-dataset training for ensemble models
([#1401](#1401))
([6a98583](6a98583))
* **models:** Extend halo exchange to non-symmetric case
([#1418](#1418))
([2037e6f](2037e6f))
* **models:** Refuse dropout in processors and mappers when the model is
sharded ([#1497](#1497))
([666af83](666af83))
* **models:** Remove unused layer norms from TransformerMapperBlock
([#1443](#1443))
([8c3ab41](8c3ab41))
* **models:** Replace the Triton import check in gt.py
([#1409](#1409))
([fe77248](fe77248))
* **models:** Skip input-only datasets in predict_step post-processing
([#1453](#1453))
([dd9c266](dd9c266))
* **models:** Spatial_pre_processors attribute access
([#1410](#1410))
([b52b034](b52b034))
* **models:** Subset forcings using regular indexing instead of
torch.index_select to fix device mismatch
([#1429](#1429))
([d09c61d](d09c61d))
* **tests:** Make tests discoverable
([#1359](#1359))
([54db8ce](54db8ce))
* Torch2.6 compatibility
([#1383](#1383))
([a48ce21](a48ce21))
* **training,models,graphs:** Bump torch version 2.11
([#1455](#1455))
([e21aefd](e21aefd))


### Performance Improvements

* **models,training:** Triton-transformer-backend
([#1444](#1444))
([6636d69](6636d69))
</details>

---
> [!IMPORTANT]
> Please do not change the PR title, manifest file, or any other
automatically generated content in this PR unless you understand the
implications. Changes here can break the release process.
> 
> ⚠️ Merging this PR will:
> - Create a new release
> - Trigger deployment pipelines
> - Update package versions

 **Before merging:**
 - Ensure all tests pass
 - Review the changelog carefully
 - Get required approvals

[Release-please
documentation](https://github.com/googleapis/release-please)

Nov 3, 2026
6a2377d
zip
tar.gz
Notes

training-0.17.1

chore: Release hotfix/dataset-join-scalar (#1437)

<details><summary>training: 0.17.1</summary>

##
[0.17.1](training-0.17.0...training-0.17.1)
(2026-10-01)


### Bug Fixes

* Correct tendency scaler with dataset join
([3ebeac3](3ebeac3))
</details>

<details><summary>graphs: 0.9.8</summary>

##
[0.9.8](graphs-0.9.7...graphs-0.9.8)
(2026-10-01)


### Bug Fixes

* **graphs:** Pin xarray
([#1436](#1436))
([3681ce7](3681ce7))
</details>

Oct 1, 2026
984d0d4
zip
tar.gz
Notes

graphs-0.9.8

chore: Release hotfix/dataset-join-scalar (#1437)

<details><summary>training: 0.17.1</summary>

##
[0.17.1](training-0.17.0...training-0.17.1)
(2026-10-01)


### Bug Fixes

* Correct tendency scaler with dataset join
([3ebeac3](3ebeac3))
</details>

<details><summary>graphs: 0.9.8</summary>

##
[0.9.8](graphs-0.9.7...graphs-0.9.8)
(2026-10-01)


### Bug Fixes

* **graphs:** Pin xarray
([#1436](#1436))
([3681ce7](3681ce7))
</details>

Oct 1, 2026
984d0d4
zip
tar.gz
Notes

training-0.17.0

chore: Release main (#1323)

<details><summary>training: 0.17.0</summary>

##
[0.17.0](training-0.16.0...training-0.17.0)
(2026-09-07)


### ⚠ BREAKING CHANGES

* **configs:** expand configurability for models with multiple
encoders/decoders
([#1229](#1229))

### Features

* **configs:** Expand configurability for models with multiple
encoders/decoders
([#1229](#1229))
([98a16a0](98a16a0))
* **training:** Flexible timestepping for forecaster
([#1070](#1070))
([fdeb9cf](fdeb9cf))
* **training:** Graph scores
([#1268](#1268))
([11e8498](11e8498))
* **training:** Strategy distribute by 1
([#1346](#1346))
([3bf1a70](3bf1a70))


### Bug Fixes

* Circular import
([#1322](#1322))
([89036bc](89036bc))
* **documentation, training:** Clean-up of graph scores and fix
documentation
([#1338](#1338))
([5aa6585](5aa6585))
* **training:** Honour diagnostics.enable_checkpointing in the trainer
([#1332](#1332))
([2236216](2236216))
* **training:** Restore rollout and dataloader epoch on restart
([#1339](#1339))
([c6d79c1](c6d79c1))


### Performance Improvements

* **training:** Use tensor.sum() on the scaling mask in
ReweightedGraphNodeAttributeScaler
([#1328](#1328))
([ce8d0d6](ce8d0d6))


### Reverts

* **training:** "feat: reuse processing callbacks
([#1213](#1213))"
([#1319](#1319))
([93476b7](93476b7))
</details>

<details><summary>graphs: 0.9.7</summary>

##
[0.9.7](graphs-0.9.6...graphs-0.9.7)
(2026-09-07)


### Bug Fixes

* **graphs:** Correct planar area weights at the domain edge
([#1320](#1320))
([d55d7e7](d55d7e7))
* **graphs:** Use the correct CUDA device when building graphs on
multiple GPUs
([#1333](#1333))
([e109ce8](e109ce8))
</details>

<details><summary>models: 0.19.0</summary>

##
[0.19.0](models-0.18.0...models-0.19.0)
(2026-09-07)


### ⚠ BREAKING CHANGES

* **configs:** expand configurability for models with multiple
encoders/decoders
([#1229](#1229))

### Features

* **configs:** Expand configurability for models with multiple
encoders/decoders
([#1229](#1229))
([98a16a0](98a16a0))
* **training:** Graph scores
([#1268](#1268))
([11e8498](11e8498))


### Bug Fixes

* **documentation, training:** Clean-up of graph scores and fix
documentation
([#1338](#1338))
([5aa6585](5aa6585))


### Performance Improvements

* **models:** Keep torch off the anemoi-training CLI import path
([#1329](#1329))
([c0cc4ba](c0cc4ba))
</details>

Sep 7, 2026
c6bdcdb
zip
tar.gz
Notes

models-0.19.0

chore: Release main (#1323)

<details><summary>training: 0.17.0</summary>

##
[0.17.0](training-0.16.0...training-0.17.0)
(2026-09-07)


### ⚠ BREAKING CHANGES

* **configs:** expand configurability for models with multiple
encoders/decoders
([#1229](#1229))

### Features

* **configs:** Expand configurability for models with multiple
encoders/decoders
([#1229](#1229))
([98a16a0](98a16a0))
* **training:** Flexible timestepping for forecaster
([#1070](#1070))
([fdeb9cf](fdeb9cf))
* **training:** Graph scores
([#1268](#1268))
([11e8498](11e8498))
* **training:** Strategy distribute by 1
([#1346](#1346))
([3bf1a70](3bf1a70))


### Bug Fixes

* Circular import
([#1322](#1322))
([89036bc](89036bc))
* **documentation, training:** Clean-up of graph scores and fix
documentation
([#1338](#1338))
([5aa6585](5aa6585))
* **training:** Honour diagnostics.enable_checkpointing in the trainer
([#1332](#1332))
([2236216](2236216))
* **training:** Restore rollout and dataloader epoch on restart
([#1339](#1339))
([c6d79c1](c6d79c1))


### Performance Improvements

* **training:** Use tensor.sum() on the scaling mask in
ReweightedGraphNodeAttributeScaler
([#1328](#1328))
([ce8d0d6](ce8d0d6))


### Reverts

* **training:** "feat: reuse processing callbacks
([#1213](#1213))"
([#1319](#1319))
([93476b7](93476b7))
</details>

<details><summary>graphs: 0.9.7</summary>

##
[0.9.7](graphs-0.9.6...graphs-0.9.7)
(2026-09-07)


### Bug Fixes

* **graphs:** Correct planar area weights at the domain edge
([#1320](#1320))
([d55d7e7](d55d7e7))
* **graphs:** Use the correct CUDA device when building graphs on
multiple GPUs
([#1333](#1333))
([e109ce8](e109ce8))
</details>

<details><summary>models: 0.19.0</summary>

##
[0.19.0](models-0.18.0...models-0.19.0)
(2026-09-07)


### ⚠ BREAKING CHANGES

* **configs:** expand configurability for models with multiple
encoders/decoders
([#1229](#1229))

### Features

* **configs:** Expand configurability for models with multiple
encoders/decoders
([#1229](#1229))
([98a16a0](98a16a0))
* **training:** Graph scores
([#1268](#1268))
([11e8498](11e8498))


### Bug Fixes

* **documentation, training:** Clean-up of graph scores and fix
documentation
([#1338](#1338))
([5aa6585](5aa6585))


### Performance Improvements

* **models:** Keep torch off the anemoi-training CLI import path
([#1329](#1329))
([c0cc4ba](c0cc4ba))
</details>

Sep 7, 2026
c6bdcdb
zip
tar.gz
Notes

graphs-0.9.7

chore: Release main (#1323)

<details><summary>training: 0.17.0</summary>

##
[0.17.0](training-0.16.0...training-0.17.0)
(2026-09-07)


### ⚠ BREAKING CHANGES

* **configs:** expand configurability for models with multiple
encoders/decoders
([#1229](#1229))

### Features

* **configs:** Expand configurability for models with multiple
encoders/decoders
([#1229](#1229))
([98a16a0](98a16a0))
* **training:** Flexible timestepping for forecaster
([#1070](#1070))
([fdeb9cf](fdeb9cf))
* **training:** Graph scores
([#1268](#1268))
([11e8498](11e8498))
* **training:** Strategy distribute by 1
([#1346](#1346))
([3bf1a70](3bf1a70))


### Bug Fixes

* Circular import
([#1322](#1322))
([89036bc](89036bc))
* **documentation, training:** Clean-up of graph scores and fix
documentation
([#1338](#1338))
([5aa6585](5aa6585))
* **training:** Honour diagnostics.enable_checkpointing in the trainer
([#1332](#1332))
([2236216](2236216))
* **training:** Restore rollout and dataloader epoch on restart
([#1339](#1339))
([c6d79c1](c6d79c1))


### Performance Improvements

* **training:** Use tensor.sum() on the scaling mask in
ReweightedGraphNodeAttributeScaler
([#1328](#1328))
([ce8d0d6](ce8d0d6))


### Reverts

* **training:** "feat: reuse processing callbacks
([#1213](#1213))"
([#1319](#1319))
([93476b7](93476b7))
</details>

<details><summary>graphs: 0.9.7</summary>

##
[0.9.7](graphs-0.9.6...graphs-0.9.7)
(2026-09-07)


### Bug Fixes

* **graphs:** Correct planar area weights at the domain edge
([#1320](#1320))
([d55d7e7](d55d7e7))
* **graphs:** Use the correct CUDA device when building graphs on
multiple GPUs
([#1333](#1333))
([e109ce8](e109ce8))
</details>

<details><summary>models: 0.19.0</summary>

##
[0.19.0](models-0.18.0...models-0.19.0)
(2026-09-07)


### ⚠ BREAKING CHANGES

* **configs:** expand configurability for models with multiple
encoders/decoders
([#1229](#1229))

### Features

* **configs:** Expand configurability for models with multiple
encoders/decoders
([#1229](#1229))
([98a16a0](98a16a0))
* **training:** Graph scores
([#1268](#1268))
([11e8498](11e8498))


### Bug Fixes

* **documentation, training:** Clean-up of graph scores and fix
documentation
([#1338](#1338))
([5aa6585](5aa6585))


### Performance Improvements

* **models:** Keep torch off the anemoi-training CLI import path
([#1329](#1329))
([c0cc4ba](c0cc4ba))
</details>

Sep 7, 2026
c6bdcdb
zip
tar.gz
Notes

training-0.16.0

chore: Release main (#1257)

🤖 Automated Release PR

This PR was created by `release-please` to prepare the next release.
Once merged:

1. A new version tag will be created
2. A GitHub release will be published
3. The changelog will be updated

Changes to be included in the next release:
---


<details><summary>training: 0.16.0</summary>

##
[0.16.0](training-0.15.0...training-0.16.0)
(2026-08-11)


### ⚠ BREAKING CHANGES

* **training:** aggregate multiscale loss outputs
([#1265](#1265))
* **models,training:** universal loss scalers for spectral losses
([#1253](#1253))
* refactor callback interface to introduce pluggable system and
evaluation module
([#1241](#1241))

### Features

* Create and ship py.typed
([#1249](#1249))
([d0cfe4c](d0cfe4c))
* Log accuracy testing to testing server
([#1256](#1256))
([e38e2c5](e38e2c5))
* **models,training:** Universal loss scalers for spectral losses
([#1253](#1253))
([034271e](034271e))
* Refactor callback interface to introduce pluggable system and
evaluation module
([#1241](#1241))
([d456f98](d456f98))
* Weekly_bulletin
([#1260](#1260))
([7df32d5](7df32d5))


### Bug Fixes

* Forecaster and tendency statistics
([#1288](#1288))
([163e1ad](163e1ad))
* **graphs:** Use weights_only=True when loading graphs in anemoi-graphs
([#1303](#1303))
([ab292a5](ab292a5))
* Tendency temporal downscaler
([#1293](#1293))
([654f4bb](654f4bb))
* **training:** Add `base_seed` entry to metadata schema
([#1290](#1290))
([4485d82](4485d82))
* **training:** Aggregate multiscale loss outputs
([#1265](#1265))
([9e34f6a](9e34f6a))
* **training:** Clean-up after imputer fix
([#1302](#1302))
([81391ff](81391ff))
* **training:** Expandable segments warning
([#1259](#1259))
([abcaf9b](abcaf9b))
* **training:** Fix race condition in benchmark tests
([#1267](#1267))
([361e713](361e713))
* **training:** Handle NaN masks correctly with sharding
([#1283](#1283))
([d7f6233
8
](d7f6233))
* **training:** Increase checkpoint download timeout time
([#1272](#1272))
([ec0cbbe](ec0cbbe))
* **training:** Protect against seed overflow
([#1228](#1228))
([eb0f6e1](eb0f6e1))
* **training:** Remove broken n_plots_per_sample option from ensemble
plots ([#1275](#1275))
([528483f](528483f))
* **training:** Remove unused num_features from config and schemas
([#1306](#1306))
([9fe1411](9fe1411))
* **training:** Set default progress bar for profiler
([#1277](#1277))
([4dd8739](4dd8739))
* **training:** Update accuracy test after fix to seed overflow
([#1228](#1228))
([#1263](#1263))
([d8b4feb](d8b4feb))


### Performance Improvements

* **training:** Disable broadcast buffers
([#1278](#1278))
([37d28df](37d28df))
* **training:** Minimise broadcasts from lightning checkpoints
([#1279](#1279))
([1ef4546](1ef4546))
* **training:** Skip redundant `advance_input`
([#1285](#1285))
([14040fd](14040fd))


### Reverts

* "fix(graphs): use weights_only=True when loading graphs in
anemoi-graphs"
([#1311](#1311))
([c30e3e8](c30e3e8))
</details>

<details><summary>graphs: 0.9.6</summary>

##
[0.9.6](graphs-0.9.5...graphs-0.9.6)
(2026-08-11)


### Features

* Create and ship py.typed
([#1249](#1249))
([d0cfe4c](d0cfe4c))


### Bug Fixes

* **graphs:** Replace qhull joggle with merged facets in
PlanarAreaWeights
([#1298](#1298))
([80af680](80af680))
* **graphs:** Use weights_only=True when loading graphs in anemoi-graphs
([#1303](#1303))
([ab292a5](ab292a5))


### Reverts

* "fix(graphs): use weights_only=True when loading graphs in
anemoi-graphs"
([#1311](#1311))
([c30e3e8](c30e3e8))
</details>

<details><summary>models: 0.18.0</summary>

##
[0.18.0](models-0.17.0...models-0.18.0)
(2026-08-11)


### ⚠ BREAKING CHANGES

* **models,training:** universal loss scalers for spectral losses
([#1253](#1253))

### Features

* Create and ship py.typed
([#1249](#1249))
([d0cfe4c](d0cfe4c))
* **models,training:** Universal loss scalers for spectral losses
([#1253](#1253))
([034271e](034271e))
* **models:** Allow torch compile to go through triton GT
([#1271](#1271))
([ec11bf2](ec11bf2))
* **models:** Comm primitives tests
([#1234](#1234))
([21851c7](21851c7))


### Bug Fixes

* **models,tests:** Wrong expected shapes in graphtransformer test
([#1287](#1287))
([0fa84c1](0fa84c1))
* **models:** Trainable edge perm migration script
([#1294](#1294))
([e7b0ef3](e7b0ef3))
* **training:** Handle NaN masks correctly with sharding
([#1283](#1283))
([d7f6233](d7f6233))


### Performance Improvements

* **training:** All-to-all interface for torch.compile()
([#1282](#1282))
([3de62a4](3de62a4))
</details>

---
> [!IMPORTANT]
> Please do not change the PR title, manifest file, or any other
automatically generated content in this PR unless you understand the
implications. Changes here can break the release process.
> 
> ⚠️ Merging this PR will:
> - Create a new release
> - Trigger deployment pipelines
> - Update package versions

 **Before merging:**
 - Ensure all tests pass
 - Review the changelog carefully
 - Get required approvals

[Release-please
documentation](https://github.com/googleapis/release-please)

Aug 11, 2026
c3c7a89
zip
tar.gz
Notes

models-0.18.0

chore: Release main (#1257)

🤖 Automated Release PR

This PR was created by `release-please` to prepare the next release.
Once merged:

1. A new version tag will be created
2. A GitHub release will be published
3. The changelog will be updated

Changes to be included in the next release:
---


<details><summary>training: 0.16.0</summary>

##
[0.16.0](training-0.15.0...training-0.16.0)
(2026-08-11)


### ⚠ BREAKING CHANGES

* **training:** aggregate multiscale loss outputs
([#1265](#1265))
* **models,training:** universal loss scalers for spectral losses
([#1253](#1253))
* refactor callback interface to introduce pluggable system and
evaluation module
([#1241](#1241))

### Features

* Create and ship py.typed
([#1249](#1249))
([d0cfe4c](d0cfe4c))
* Log accuracy testing to testing server
([#1256](#1256))
([e38e2c5](e38e2c5))
* **models,training:** Universal loss scalers for spectral losses
([#1253](#1253))
([034271e](034271e))
* Refactor callback interface to introduce pluggable system and
evaluation module
([#1241](#1241))
([d456f98](d456f98))
* Weekly_bulletin
([#1260](#1260))
([7df32d5](7df32d5))


### Bug Fixes

* Forecaster and tendency statistics
([#1288](#1288))
([163e1ad](163e1ad))
* **graphs:** Use weights_only=True when loading graphs in anemoi-graphs
([#1303](#1303))
([ab292a5](ab292a5))
* Tendency temporal downscaler
([#1293](#1293))
([654f4bb](654f4bb))
* **training:** Add `base_seed` entry to metadata schema
([#1290](#1290))
([4485d82](4485d82))
* **training:** Aggregate multiscale loss outputs
([#1265](#1265))
([9e34f6a](9e34f6a))
* **training:** Clean-up after imputer fix
([#1302](#1302))
([81391ff](81391ff))
* **training:** Expandable segments warning
([#1259](#1259))
([abcaf9b](abcaf9b))
* **training:** Fix race condition in benchmark tests
([#1267](#1267))
([361e713](361e713))
* **training:** Handle NaN masks correctly with sharding
([#1283](#1283))
([d7f6233](d7f6233))
* **training:** Increase checkpoint download timeout time
([#1272](#1272))
([ec0cbbe](ec0cbbe))
* **training:** Protect against seed overflow
([#1228](#1228))
([eb0f6e1](eb0f6e1))
* **training:** Remove broken n_plots_per_sample option from ensemble
plots ([#1275](#1275))
([528483f](528483f))
* **training:** Remove unused num_features from config and schemas
([#1306](#1306))
([9fe1411](9fe1411))
* **training:** Set default progress bar for profiler
([#1277](#1277))
([4dd8739](4dd8739))
* **training:** Update accuracy test after fix to seed overflow
([#1228](#1228))
([#1263](#1263))
([d8b4feb](d8b4feb))


### Performance Improvements

* **training:** Disable broadcast buffers
([#1278](#1278))
([37d28df](37d28df))
* **training:** Minimise broadcasts from lightning checkpoints
([#1279](#1279))
([1ef4546](1ef4546))
* **training:** Skip redundant `advance_input`
([#1285](#1285))
([14040fd](14040fd))


### Reverts

* "fix(graphs): use weights_only=True when loading graphs in
anemoi-graphs"
([#1311](#1311))
([c30e3e8](c30e3e8))
</details>

<details><summary>graphs: 0.9.6</summary>

##
[0.9.6](graphs-0.9.5...graphs-0.9.6)
(2026-08-11)


### Features

* Create and ship py.typed
([#1249](#1249))
([d0cfe4c](d0cfe4c))


### Bug Fixes

* **graphs:** Replace qhull joggle with merged facets in
PlanarAreaWeights
([#1298](#1298))
([80af680](80af680))
* **graphs:** Use weights_only=True when loading graphs in anemoi-graphs
([#1303](#1303))
([ab292a5](ab292a5))


### Reverts

* "fix(graphs): use weights_only=True when loading graphs in
anemoi-graphs"
([#1311](#1311))
([c30e3e8](c30e3e8))
</details>

<details><summary>models: 0.18.0</summary>

##
[0.18.0](models-0.17.0...models-0.18.0)
(2026-08-11)


### ⚠ BREAKING CHANGES

* **models,training:** universal loss scalers for spectral losses
([#1253](#1253))

### Features

* Create and ship py.typed
([#1249](#1249))
([d0cfe4c](d0cfe4c))
* **models,training:** Universal loss scalers for spectral losses
([#1253](#1253))
([034271e](034271e))
* **models:** Allow torch compile to go through triton GT
([#1271](#1271))
([ec11bf2](ec11bf2))
* **models:** Comm primitives tests
([#1234](#1234))
([21851c7](21851c7))


### Bug Fixes

* **models,tests:** Wrong expected shapes in graphtransformer test
([#1287](#1287))
([0fa84c1](0fa84c1))
* **models:** Trainable edge perm migration script
([#1294](#1294))
([e7b0ef3](e7b0ef3))
* **training:** Handle NaN masks correctly with sharding
([#1283](#1283))
([d7f6233](d7f6233))


### Performance Improvements

* **training:** All-to-all interface for torch.compile()
([#1282](#1282))
([3de62a4](3de62a4))
</details>

---
> [!IMPORTANT]
> Please do not change the PR title, manifest file, or any other
automatically generated content in this PR unless you understand the
implications. Changes here can break the release process.
> 
> ⚠️ Merging this PR will:
> - Create a new release
> - Trigger deployment pipelines
> - Update package versions

 **Before merging:**
 - Ensure all tests pass
 - Review the changelog carefully
 - Get required approvals

[Release-please
documentation](https://github.com/googleapis/release-please)

Aug 11, 2026
c3c7a89
zip
tar.gz
Notes

PreviousNext