Bump the all-python-packages group across 1 directory with 22 updates

[dependabot]

Bumps the all-python-packages group with 18 updates in the / directory:

Package	From	To
accelerate	1.10.1	1.12.0
flwr	1.22.0	1.25.0
pyarrow	21.0.0	23.0.0
pydantic	2.12.0	2.12.5
pydantic-settings	2.11.0	2.12.0
torch	2.8.0	2.9.1
peft	0.17.1	0.18.1
transformers	4.56.2	4.57.6
sentence-transformers	5.1.1	5.2.0
grpcio	1.75.1	1.76.0
llama-index-core	0.14.4	0.14.12
langchain-core	0.3.79	1.2.7
mypy	1.18.2	1.19.1
pre-commit	4.3.0	4.5.1
pylint	4.0.0	4.0.4
pytest	8.4.2	9.0.2
ruff	0.14.0	0.14.13
mkdocs-material	9.6.21	9.7.1

Updates accelerate from 1.10.1 to 1.12.0

Release notes

Sourced from accelerate's releases.

v1.12.0: Deepspeed Ulysses/ALST

Deepspeed Ulysses/ALST integration

Deepspeed Ulysses/ALST is an efficient way of training on long sequences by employing sequence parallelism and attention head parallelism. You can learn more about this technology in this paper https://arxiv.org/abs/2506.13996 or this deepspeed tutorial https://www.deepspeed.ai/tutorials/ulysses-alst-sequence-parallelism/.

To enable Deepspeed Ulysses, you first need to create ParallelismConfig and setting sp related args:

parallelism_config = ParallelismConfig(
    sp_backend="deepspeed",
    sp_size=2,
    sp_handler=DeepSpeedSequenceParallelConfig(...),
)

Then, you need to make sure to compute the correct loss as described on our docs

        ...
        losses_per_rank = torch.distributed.nn.functional.all_gather(loss, group=sp_group)
        good_tokens = (shift_labels != -100).view(-1).sum()
        good_tokens_per_rank = torch.distributed.nn.functional.all_gather(good_tokens, group=sp_group)
        total_loss = sum(
            losses_per_rank[rank] * good_tokens_per_rank[rank]
            for rank in range(sp_world_size)
            if good_tokens_per_rank[rank] > 0
        )
        total_good_tokens = sum(good_tokens_per_rank)
        loss = total_loss / max(total_good_tokens, 1)

Thanks @​S1ro1 for starting this work and for @​stas00 for finishing this work. Also thanks @​kashif for adding docs and reviewing/testing this PR !

This feature will also be available in HF Trainer thanks for this PR from @​stas00: huggingface/transformers#41832

Minor changes

• Remove warning for cpu_ram_efficient_loading by @​SunMarc in huggingface/accelerate#3816
• update typo in bnb quantisation 4bit flag docstring by @​hbraith in huggingface/accelerate#3828
• ArXiv -> HF Papers by @​qgallouedec in huggingface/accelerate#3834
• Fix typo in broadcast_object_list docstring by @​wsntxxn in huggingface/accelerate#3823
• [Bug] Update torch.optim.Optimizer parameter states after tensor parallelism by @​naomili0924 in huggingface/accelerate#3835
• use self hosted runner by @​SunMarc in huggingface/accelerate#3841
• device type helper by @​kashif in huggingface/accelerate#3843

New Contributors

• @​hbraith made their first contribution in huggingface/accelerate#3828

... (truncated)

Commits

• 61ff833 Release: v1.12.0
• e3655b9 device type helper (#3843)
• c7e59dd Deepspeed Ulysses/ALST integration (#3817)
• e8f241f use self hosted runner (#3841)
• 00b1b18 [Bug] Update torch.optim.Optimizer parameter states after tensor parallelism ...
• a73fd3a Fix typo in broadcast_object_list docstring (#3823)
• a08fb32 Change arxiv to hf papers (#3834)
• 61bcdaa update bnb quantisation 4bit flag (#3828)
• a12beee Remove warning for cpu_ram_efficient_loading (#3816)
• 0a0c989 v1.12.0dev
• Additional commits viewable in compare view

Updates datasets from 4.2.0 to 3.1.0

Commits

• dfb52e2 Release: 3.1.0 (#7270)
• f8e3321 Disallow video push_to_hub (#7265)
• 46e4616 fix docs relative links (#7264)
• b08f762 Small addition to video docs (#7263)
• 1946182 Allow video with disabeld decoding without decord (#7262)
• f75f489 Don't embed videos (#7259)
• ff0149f Always set non-null writer batch size (#7258)
• 444ce83 Retry all requests timeouts (#7256)
• a9c610f fix ci for pyarrow 18 (#7257)
• 91e9313 fix decord import (#7255)
• Additional commits viewable in compare view

Updates flwr from 1.22.0 to 1.25.0

Release notes

Sourced from flwr's releases.

Flower 1.25.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Chong Shen Ng, Daniel J. Beutel, Heng Pan, Javier, Mohammad Naseri, Soumik Sarker, William Lindskog, Yan Gao, sarahhfalco

What's new?

• Track compute time and network traffic for runs (#6241, #6242, #6243, #6244, #6245, #6249, #6266, #6267, #6268, #6269, #6270, #6271, #6272, #6273, #6274, #6275, #6276, #6279)

  Flower now records compute time and network traffic for a run. The run detail view shown by flwr list --run-id <run-id> displays traffic exchanged between SuperLink and SuperNode, as well as compute time used by ServerApp and ClientApp.

• Refactor flwr new to pull apps from the Flower platform (#6251, #6252, #6258, #6259, #6260, #6263, #6265, #6283, #6284, #6285, #6292, #6294, #6302)

  Refactors flwr new to fetch Flower apps directly from the Flower platform (see the usage reference). This introduces new and updated quickstart examples (including NumPy and FlowerTune LLM), renames and updates existing examples, aligns CI to run against platform-backed examples, and updates related documentation and benchmark instructions.

• Migrate examples to the Message API and remove outdated Docker Compose as well as Tensorflow Privacy examples (#6232, #6233, #6238, #6297, #6304)

  Migrates the scikit-learn, Vertical FL, and Opacus examples to the Message API, with the Vertical FL example also updated to use flwr-datasets. The outdated Docker Compose and Tensorflow Privacy examples are removed.

• Improve CLI output with human-readable durations (#6277, #6296)

  Updates the Flower CLI to display durations in a more human-friendly format (xd xh xm xs), automatically selecting appropriate units instead of the previous HH:MM:SS format.

• Update examples and baselines (#6234, #6256, #6257, #6264, #6280, #6281, #6286, #6287, #6288, #6290, #6291, #6293)

• Improve documentation (#6229, #6230, #6255, #6262)

• Update CI/CD configuration (#6168, #6246, #6295, #6305)

• General improvements (#6056, #6085, #6176, #6235, #6236, #6254, #6278, #6299)

  As always, many parts of the Flower framework and quality infrastructure were improved and updated.

Incompatible changes

• Remove bundled templates from flwr new (#6261)

  Removes the templates previously bundled with the Flower wheel now that flwr new pulls apps from the Flower platform. The --framework and --username options are deprecated as part of this change.

Flower 1.24.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Charles Beauville, Chong Shen Ng, Daniel J. Beutel, Daniel Nata Nugraha, Heng Pan, Javier, Patrick Foley, Robert Steiner, Yan Gao

What's new?

... (truncated)

Commits

• a66982d feat(framework): Update changelog for v1.25.0 (#6298)
• e3d1aa3 Update text and language files (#6306)
• 7d560bc refactor(examples): Remove tensorflow-privacy example (#6304)
• 533a511 docs(examples): Update jax tutorial based on the app `@flwrlabs/quickstart-...
• adc9239 ci(*:skip): Fix heartbeat CI test (#6305)
• 7156052 Update text and language files (#6303)
• b3c81d5 docs(framework): Update flwr new steps in the tutorials (#6283)
• 6cf2800 refactor(examples): Update opacus app (#6297)
• 547c93e docs(framework): Update flwr new instances in documentation (#6294)
• a918b51 build(deps): Bump next in /intelligence/ts/examples/nextjs-web-chats (#6300)
• Additional commits viewable in compare view

Updates pyarrow from 21.0.0 to 23.0.0

Release notes

Sourced from pyarrow's releases.

Apache Arrow 23.0.0

Release Notes URL: https://arrow.apache.org/release/23.0.0.html

Apache Arrow 23.0.0 RC2

Release Notes: Release Candidate: 23.0.0 RC2

Apache Arrow 22.0.0

Release Notes URL: https://arrow.apache.org/release/22.0.0.html

Apache Arrow 22.0.0 RC1

Release Notes: Release Candidate: 22.0.0 RC1

Apache Arrow 22.0.0 RC0

Release Notes: Release Candidate: 22.0.0 RC0

Commits

• eafe3a9 MINOR: [Release] Update versions for 23.0.0
• 45781e8 MINOR: [Release] Update .deb/.rpm changelogs for 23.0.0
• 365e5ac MINOR: [Release] Update CHANGELOG.md for 23.0.0
• e33512d GH-48841: [Release][Package] Add GH_TOKEN to rake build step on Linux Packagi...
• 86d28e9 GH-48838: [Release] Use gh cli to download sources for Linux packages and pub...
• b6362d0 GH-48582: [CI][GPU][C++][Python] Add new CUDA jobs using the new self-hosted ...
• 8fb7a53 GH-48827: [CI][Python] Add required xz dependency to emscripten dockerfile (#...
• 342c5d5 GH-48809: [CI] Fix homebrew-cpp with Mac by using formula-based dependency re...
• 8955bbf GH-48631: [R] Non-API calls: 'ATTRIB', 'SET_ATTRIB' (#48634)
• b90a2b8 GH-48314: [Python] Compat with pandas 3.0 changed default datetime unit (#48319)
• Additional commits viewable in compare view

Updates pydantic from 2.12.0 to 2.12.5

Release notes

Sourced from pydantic's releases.

v2.12.5 2025-11-26

v2.12.5 (2025-11-26)

This is the fifth 2.12 patch release, addressing an issue with the MISSING sentinel and providing several documentation improvements.

The next 2.13 minor release will be published in a couple weeks, and will include a new polymorphic serialization feature addressing the remaining unexpected changes to the serialize as any behavior.

• Fix pickle error when using model_construct() on a model with MISSING as a default value by @​ornariece in #12522.
• Several updates to the documentation by @​Viicos.

Full Changelog: pydantic/pydantic@v2.12.4...v2.12.5

v2.12.4 2025-11-05

v2.12.4 (2025-11-05)

This is the fourth 2.12 patch release, fixing more regressions, and reverting a change in the build() method of the AnyUrl and Dsn types.

This patch release also fixes an issue with the serialization of IP address types, when serialize_as_any is used. The next patch release will try to address the remaining issues with serialize as any behavior by introducing a new polymorphic serialization feature, that should be used in most cases in place of serialize as any.

• Fix issue with forward references in parent TypedDict classes by @​Viicos in #12427.

  This issue is only relevant on Python 3.14 and greater.

• Exclude fields with exclude_if from JSON Schema required fields by @​Viicos in #12430

• Revert URL percent-encoding of credentials in the build() method of the AnyUrl and Dsn types by @​davidhewitt in pydantic-core#1833.

  This was initially considered as a bugfix, but caused regressions and as such was fully reverted. The next release will include an opt-in option to percent-encode components of the URL.

• Add type inference for IP address types by @​davidhewitt in pydantic-core#1868.

  The 2.12 changes to the serialize_as_any behavior made it so that IP address types could not properly serialize to JSON.

• Avoid getting default values from defaultdict by @​davidhewitt in pydantic-core#1853.

  This fixes a subtle regression in the validation behavior of the collections.defaultdict type.

• Fix issue with field serializers on nested typed dictionaries by @​davidhewitt in pydantic-core#1879.

• Add more pydantic-core builds for the three-threaded version of Python 3.14 by @​davidhewitt in pydantic-core#1864.

Full Changelog: pydantic/pydantic@v2.12.3...v2.12.4

v2.12.3 2025-10-17

v2.12.3 (2025-10-17)

What's Changed

This is the third 2.13 patch release, fixing issues related to the FieldInfo class, and reverting a change to the supported after model validator function signatures.

... (truncated)

Changelog

Sourced from pydantic's changelog.

v2.12.5 (2025-11-26)

GitHub release

This is the fifth 2.12 patch release, addressing an issue with the MISSING sentinel and providing several documentation improvements.

The next 2.13 minor release will be published in a couple weeks, and will include a new polymorphic serialization feature addressing the remaining unexpected changes to the serialize as any behavior.

• Fix pickle error when using model_construct() on a model with MISSING as a default value by @​ornariece in #12522.
• Several updates to the documentation by @​Viicos.

v2.12.4 (2025-11-05)

GitHub release

This is the fourth 2.12 patch release, fixing more regressions, and reverting a change in the build() method of the AnyUrl and Dsn types.

This patch release also fixes an issue with the serialization of IP address types, when serialize_as_any is used. The next patch release will try to address the remaining issues with serialize as any behavior by introducing a new polymorphic serialization feature, that should be used in most cases in place of serialize as any.

• Fix issue with forward references in parent TypedDict classes by @​Viicos in #12427.

  This issue is only relevant on Python 3.14 and greater.

• Exclude fields with exclude_if from JSON Schema required fields by @​Viicos in #12430

• Revert URL percent-encoding of credentials in the build() method of the AnyUrl and Dsn types by @​davidhewitt in pydantic-core#1833.

  This was initially considered as a bugfix, but caused regressions and as such was fully reverted. The next release will include an opt-in option to percent-encode components of the URL.

• Add type inference for IP address types by @​davidhewitt in pydantic-core#1868.

  The 2.12 changes to the serialize_as_any behavior made it so that IP address types could not properly serialize to JSON.

• Avoid getting default values from defaultdict by @​davidhewitt in pydantic-core#1853.

  This fixes a subtle regression in the validation behavior of the collections.defaultdict type.

• Fix issue with field serializers on nested typed dictionaries by @​davidhewitt in pydantic-core#1879.

• Add more pydantic-core builds for the three-threaded version of Python 3.14 by @​davidhewitt in pydantic-core#1864.

v2.12.3 (2025-10-17)

GitHub release

... (truncated)

Commits

• bd2d0dd Prepare release v2.12.5
• 7d0302e Document security implications when using create_model()
• e9ef980 Fix typo in Standard Library Types documentation
• f2c20c0 Add pydantic-docs dev dependency, make use of versioning blocks
• a76c1aa Update documentation about JSON Schema
• 8cbc72c Add documentation about custom __init__()
• 99eba59 Add additional test for FieldInfo.get_default()
• c710769 Special case MISSING sentinel in smart_deepcopy()
• 20a9d77 Do not delete mock validator/serializer in rebuild_dataclass()
• c86515a Update parts of the model and revalidate_instances documentation
• Additional commits viewable in compare view

Updates pydantic-settings from 2.11.0 to 2.12.0

Release notes

Sourced from pydantic-settings's releases.

v2.12.0

What's Changed

• Support for enum kebab case. by @​kschwab in pydantic/pydantic-settings#686
• Apply source order: init > env > dotenv > secrets > defaults and pres… by @​chbndrhnns in pydantic/pydantic-settings#688
• Add NestedSecretsSettings source by @​makukha in pydantic/pydantic-settings#690
• Strip non-explicit default values. by @​kschwab in pydantic/pydantic-settings#692
• Coerce env vars if strict is True. by @​kschwab in pydantic/pydantic-settings#693
• Restore init kwarg names before returning final state dictionary. by @​kschwab in pydantic/pydantic-settings#700
• Drop Python3.9 support by @​hramezani in pydantic/pydantic-settings#699
• Adapt test_protected_namespace_defaults for dev. Pydantic by @​musicinmybrain in pydantic/pydantic-settings#637
• Add Python 3.14 by @​hramezani in pydantic/pydantic-settings#704
• Prepare release 2.12 by @​hramezani in pydantic/pydantic-settings#705

New Contributors

• @​chbndrhnns made their first contribution in pydantic/pydantic-settings#688

Full Changelog: pydantic/pydantic-settings@v2.11.0...v2.12.0

Commits

• 584983d Prepare release 2.12 (#705)
• 6b4d87e Add Python 3.14 (#704)
• 02de5b6 Adapt test_protected_namespace_defaults for dev. Pydantic (#637)
• 4239ea4 Drop Python3.9 support (#699)
• 5008c69 Restore init kwarg names before returning final state dictionary. (#700)
• 4433101 Coerce env vars if strict is True. (#693)
• 4d2ebfd Strip non-explicit default values. (#692)
• 4a6ffca Add NestedSecretsSettings source (#690)
• 7a6e96e Apply source order: init > env > dotenv > secrets > defaults and pres… (#688)
• 68563ed Support for enum kebab case. (#686)
• See full diff in compare view

Updates torch from 2.8.0 to 2.9.1

Release notes

Sourced from torch's releases.

PyTorch 2.9.1 Release, bug fix release

This release is meant to fix the following issues (regressions / silent correctness):

Tracked Regressions

Significant Memory Regression in F.conv3d with bfloat16 Inputs in PyTorch 2.9.0 (#166643) This release provides work around this issue. If you are impacted please install nvidia-cudnn package version 9.15+ from pypi. (#166480) (#167111)

Torch.compile

Fix Inductor bug when compiling Gemma (#165601) Fix InternalTorchDynamoError in bytecode_transformation (#166036) Fix silent correctness error_on_graph_break bug where non-empty checkpoint results in unwanted graph break resumption (#166586) Improve performance by avoiding recompilation with mark_static_address with cudagraphs (#162208) Improve performance by caching get_free_symbol_uses in torch inductor (#166338) Fix fix registration design for inductor graph partition for vLLM (#166458) (#165815) (#165514) Fix warning spamming in torch.compile (#166993) Fix exception related to uninitialized tracer_output variable (#163169) Fix crash in torch.bmm and torch.compile with PyTorch release 2.9.0 (#166457)

Other

Fix warning spamming on new APIs to control TF32 behavior (#166956) Fix distributed crash with non-contiguous gather inputs (#166181) Fix indexing on large tensor causes invalid configuration argument (#166974) Fix numeric issue in CUDNN_ATTENTION (#166912) (#166570) Fix symmetric memory issue with fused_scaled_matmul_reduce_scatter (#165086) Improve libtorch stable ABI documentation (#163899) Fix image display on pypi project description section (#166404)

2.9 Release Notes

PyTorch 2.9.0 Release Notes

• Highlights
• Backwards Incompatible Changes
• Deprecations
• New Features
• Improvements
• Bug fixes
• Performance
• Documentation
• Developers
• Security

Highlights

... (truncated)

Commits

• d38164a Enable Doc builds for: Minor Releases RCs. Minor and Patch Releases final RC ...
• b002562 Add doc for Symmetric Memory (#167477)
• 5811a8d [cuDNN][SDPA][Convolution] Expose cuDNN runtime version in CUDA hooks (#167327)
• f36c764 [dynamo][ez] Initialize tracer_output to None by default. (#167366)
• 6877288 Change forkserver test to only run below 3.13.8 (#167361)
• 9976b77 Cherry-pick LibTorch Stable ABI documentation (#167112 #166661 #163899) (#167...
• e6bcbbe [Inductor] No longer throw error in bmm out_dtype lowering due to tem… (#166922)
• 8f658d7 don't produce invalid grid configs (#166973) (#167158)
• 3d27d95 [GraphPartition] cache get_free_symbol_uses (#166338) (#166994)
• a06141f Delete deprecated fp32 precision warnings (#166956) (#166998)
• Additional commits viewable in compare view

Updates torchvision from 0.23.0 to 0.24.1

Release notes

Sourced from torchvision's releases.

TorchVision 0.24.1 Release

This is a patch release, which is compatible with PyTorch 2.9.1. There are no new features added.

Torchvision 0.24 release

Improving KeyPoints and Rotated boxes support!

We are releasing a tutorial on how to use KeyPoint transformations in our Transforms on KeyPoints with a preview below!

[!NOTE] These features are still in BETA status. The API are unlikely to change, but we may have some rough edges and we may make some slight bug fixes in future releases. Please let us know if you encounter any issue!

Detailed changes

Improvements

[ops] Improve efficiency of the box_area and box_iou functions by eliminating the intermediate to "xyxy" conversion (#8992) [ops] Update box operations to support arbitrary batch dimensions (#9058) [utils] Add control for the background color of label text boxes (#9204) [transforms] Add support for uint8 image format to the GaussianNoise transform (#9169) [transforms] Accelerate the resize transform on machines with AVX512 (#9190) [transforms] Better error handling in RandomApply for empty list of transforms (#9130) [documentation] New tutorial for KeyPoints transforms (#9209) [documentation] Various documentation improvements (#9186, #9180, #9172) [code quality] Various code quality improvements (#9193, #9161, #9201, #9218, #9160)

Bug Fixes and deprecations

[transforms] Fix output of some geometric transforms for rotated boxes (#9181, #9175) [transforms] Fix clamping for key points and add sanitization feature (#9236, #9235) [datasets] Update download links to official repo for the Caltech-101 & 256 datasets (#9205) [ops] Raise error in drop_block[2,3]d by enforcing odd-sized block sizes (#9157) [io] Removed deprecated video_reader video decoding backend. (#9208)

Contributors

🎉 We're grateful for our community, which helps us improve Torchvision by submitting issues and PRs, and providing feedback and suggestions. The following persons have contributed patches for this release: @​alperenunlu, @​AndreiMoraru123, @​atalman, @​AntoineSimoulin, @​5had3z, @​dcasbol, @​GdoongMathew, @​hrsvrn, @​JonasKlotz, @​zklaus, @​NicolasHug, @​rdong8, @​scotts, @​get9, @​diaz-esparza, @​ZainRizvi, @​Callidior, and @​pytorch/xla-devs

Commits

• d801a34 [release-only] bump version to 0.24.1
• 7a9db90 Update min version to 3.10 for release 2.9 (#9239) (#9240)
• 0429d73 [release/0.24] cherrypicks for keypoints fix (#9238)
• b919bd0 [release/0.24] cherrypicks for windows failing jobs (#9230)
• 973f5cd [release/0.24] Cherry pick (#9217)
• e437e35 [release/0.24] update version (#9214)
• fa174c6 [release/0.24] Remove prototype (#9211)
• 13b32e2 [release/0.24] Update jobs branch (#9212)
• 15f9128 Skeleton for keypoints tutorial (#9209)
• a8dc530 Don't build video decoder homemade backend by default (#9208)
• Additional commits viewable in compare view

Updates peft from 0.17.1 to 0.18.1

Release notes

Sourced from peft's releases.

0.18.1

Small patch release containing the following changes:

• #2934: Small fixes required for some special cases to work with the upcoming transformers v5 release
• #2963: Fix to enable PEFT to run with AMD ROCm thanks to @​vladmandic
• #2976: Fix a regression that inadvertently required transformers >= 4.52

0.18.0: RoAd, ALoRA, Arrow, WaveFT, DeLoRA, OSF, and more

Highlights

FIXME update list of all changes, so some more commits were added

New Methods

RoAd

@​ppetrushkov added RoAd: 2D Rotary Adaptation to PEFT in #2678. RoAd learns 2D rotation matrices that are applied using only element-wise multiplication, thus promising very fast inference with adapters in unmerged state.

Remarkably, besides LoRA, RoAd is the only PEFT method that supports mixed adapter batches. This means that when you have loaded a model with multiple RoAd adapters, you can use all of them for different samples in the same batch, which is much more efficient than switching adapters between batches:

model = PeftModel.from_pretrained(base_model, <path-to-road-adapter-A>, adapter_name="adapter-A")
model.add_adapter("adapter-B", <path-to-road-adapter-B>)
inputs = ...  # input with 3 samples
apply adapter A to sample 0, adapter B to sample 1, and use the base model for sample 2:
adapter_names = ["adapter-A", "adapter-B", "base"]
output_mixed = model(**inputs, adapter_names=adapter_names)
gen_mixed = model.generate(**inputs, adapter_names=adapter_names)

ALoRA

Activated LoRA is a technique added by @​kgreenewald in #2609 for causal language models, allowing to selectively enable LoRA adapters depending on a specific token invocation sequence in the input. This has the major benefit of being able to re-use most of the KV cache during inference when the adapter is only used to generate part of the response, after which the base model takes over again.

Arrow & GenKnowSub

@​TheTahaaa contributed not only support for Arrow, a dynamic routing algorithm between multiple loaded LoRAs in #2644, but also GenKnowSub, a technique built upon Arrow where the 'library' of LoRAs available to Arrow is first modified by subtracting general knowledge adapters (e.g., trained on subsets of Wikipedia) to enhance task-specific performance.

WaveFT

Thanks to @​Bilican, Wavelet Fine-Tuning (WaveFT) was added to PEFT in #2560. This method trains sparse updates in the wavelet domain of residual matrices, which is especially parameter efficient. It is very interesting for image generation, as it promises to generate diverse outputs while preserving subject fidelity.

DeLoRA

Decoupled Low-rank Adaptation (DeLoRA) was added by @​mwbini in #2780. This new PEFT method is similar to DoRA in so far as it decouples the angle and magnitude of the learned adapter weights. However, DeLoRA implements this in a way that promises to better prevent divergence. Moreover, it constrains the deviation of the learned weight by imposing an upper limit of the norm, which can be adjusted via the delora_lambda parameter.

OSF

... (truncated)

Commits

• e3398fc Patch release 0.18.1 (#2983)
• 77daa8d Release: v0.18.0 (#2908)
• 52e8659 Add forward compat. for tied_weights_keys dicts (#2902)
• 7d4c9fd CI: Fix issues caused by pytest v9 (#2904)
• e82e72a FIX Prefix tuning with Qwen3 issue (#2883)
• ff00848 FIX: Exploit trust_remote_code in prompt tuning (#2896)
• 0eb4668 TST Fix 4bit regression test on XPU (#2843)
• ed00f24 DOC Up...

  Description has been truncated