Repository navigation
Highlights
New methods
PaCA (Partial Connection Adaptation)
PaCA documentation (Partial Connection Adaptation, ICLR 2025), added in #3927 by @WooSunghyeon, is a PEFT method that fine-tunes randomly selected input columns ("partial connections") of the pretrained weights instead of adding adapter layers. This eliminates the sequential computation of adapter layers and stores only partial activations for the backward pass. Therefore, PaCA trains faster and uses less memory than LoRA with comparable accuracy (from the paper: 22% less training time, 16% less memory, 23% longer sequences). Use it when training time or activation memory is the bottleneck, also with quantized models (QPaCA). It is less effective when there are few tokens per step or only light modules such as q/v are adapted, and it supports only nn.Linear layers.
Astra initialization for LoRA
Astra (Activation-Space Tail-Eigenvector Low-Rank Adaptation of Large Language Models), added in by @LyoAI in #3725, uses a small task-specific calibration dataset to estimate each target layer's output-activation covariance. It initializes LoRA in the subspace of the covariance matrix's lowest-variance directions, giving the adapter a task-aware starting point. Set init_lora_weights="astra" and run preprocess_astra before injecting the adapter.
NoRA initialization for LoRA
@Joluck implemented Normalized Low-Rank Adaptation initialization for LoRA in #3908. Like normal LoRA, the B weight is initalized to zero so the adapter starts as a no-op, but each column of the randomly initialized LoRA A matrix is normalized along the rank axis. Experimentally, this resulted in faster and better conversion compared to standard LoRA initialization. Set the lora_alpha value equal to the rank r when passing init_lora_weights="nora".
Tensor Parallelism (TP) and FSDP2
One focus of the v0.22.0 release of PEFT was to improve the support for TP and FSDP2, which came through numerous changes. Please install the latest Transformers version (at least v5.17) to benefit from these improvements.
Tensor Parallelism with LoRA
LoRA now supports the embedding_colwise TP plan. Replicated LoRA weights also start from synchronized values across TP ranks, including with init_lora_weights=False and on output heads that gather their output (#4112, #3965). Please also check the updated LoRA TP guide.
FSDP2 training, checkpointing, and merging
- LoRA now infers the full layer shape on FSDP2 and HSDP meshes with custom dimension names. Switching adapters or temporarily disabling them for a reference pass preserves trainability on FSDP2 shards. Exiting
disable_adapter()reshards only when needed (#3838, #3839, #3895, #3938). - Adapter exports gather FSDP2 and HSDP adapter DTensors and correctly find adapter weights under nested FSDP or activation-checkpointing wrappers. With sharded adapter weights, call
save_pretrained()on every participating rank and setis_main_process=Trueonly on the rank that writes the files (#4065, #4108). merge_adapter()andunmerge_adapter()now update the FSDP2 parameter shards even whenreshard_after_forward=False. LoRA also preserves registeredMultiheadAttentionparameter identities during FSDP2 setup (#4077, #4083).- The FSDP guide and FSDP2 example now include an FSDP2 workflow covering training, adapter export, and resuming from a checkpoint (#4107).
Big thanks to @hivaze, @rupeshpoojary9, @Levius-Fubuki, @RerankerGuo, and @DaoyuanLi2816.
Enhancements
PEFT model health check
Call model.print_healthcheck() for a summary of adapter state, dtypes, device, etc. (or model.healthcheck() for a JSON-serializable report). It can flag issues that PEFT detects with the PEFT model. Since it also reports the number of trainable and total parameters, it should be used in place of model.print_trainable_parameters() (#3798, #3944, #4089 thanks to @eSVeeF).
Example output for a LoRA on Llama-3.2-1B (no errors were found):
Model:
ID: meta-llama/Llama-3.2-1B, type: LlamaForCausalLM, adapters: default (LORA)
Environment:
peft: 0.21.0, transformers: 5.18.0.dev0, torch: 2.14.0+cu130, Python: 3.13.9
Runtime:
training: True, devices: cpu, dtypes: bfloat16=1,235,814,400, float32=851,968, gradient checkpointing: False
Parameters:
trainable: 851,968, total: 1,236,666,368, percent trainable: 0.0689%, adapter layers: 32
State:
adapter is enabled: True, active: default, merged: none
Layer types:
default:
lora.Linear=32
LoRA hotswapping across different target layers
PEFT already supports hotswapping LoRA adapters for some time, but the first loaded adapter must be able to accomodate all following adapters. peft.helpers.create_dummy_lora_config helps with this by building a dummy starting adapter with enough rank and layer coverage for a set of compatible LoRA adapters (#3739). PEFT also now checks the adapter being replaced and resolves its rank and alpha patterns during hotswapping (#3865, #3752, thanks to @DaoyuanLi2816).
More adapters can be converted to LoRA
LoRA conversion, i.e. the conversion of other PEFT methods to LoRA adapters, now supports OFT and BOFT linear adapters, DEFT, and standard and mini MiSS modes (#3360).
Use this feature to make your PEFT checkpoints compatible with frameworks that only support LoRA.
More efficient adapter merging
LoRA and several other methods scale their merge deltas in place, avoiding an extra full-size allocation per layer. This should save a little bit of memory and should speed up merging (#4147, thanks to @Hang-get).
Changes
Python support
PEFT now requires Python 3.11 or newer, Python 3.10 has reached its end of life.
Adapter loading and target validation
PeftModel.from_pretrained()andload_adapter()now warn by default about both missing and unexpected adapter keys. Usestrict_adapter_check="error"to fail on a mismatch or"ignore"to suppress these checks (#3832 by @rupeshpoojary9).- Adapter injection validates all targets before modifying the model (#3981 by @eSVeeF).
Activation checkpointing for k-bit training
prepare_model_for_kbit_training() now uses the Transformers default for activation checkpointing (aka "gradient checkpointing") when no kwargs are supplied, selecting non-reentrant checkpointing on current Transformers versions. Pass activation_checkpointing_kwargs={“use_reentrant”: True} in the checkpointing kwargs to opt into reentrant checkpointing (#4146).
low_cpu_mem_usage
low_cpu_mem_usage=True now keeps dtypes, i.e. doesn’t restore the dtypes in the model to the ones from the checkpoint (which would most likely result in a type mismatch error, see #3509 by @vineethsaivs)
OFT
In #4157, @MaiAditya fixed the forward code of OFT Conv2d layers, which previously could produce incorrect results. This means that existing OFT checkpoints that target Conv2d layers may produce different outputs when used without merging.
modules_to_save suffix matching logic
modules_to_save used to match purely by suffix, but now the whole sub-module name must match. This is best understood with an example: if you pass modules_to_save=["classifier"], it used to match model.classifier and also model.other_classifer. Now it only matches model.classifier. If you want to keep the previous behavior, please pass modules_to_save=["classifier", "other_classifier"] or edit the adapter_config.json accordingly if you have an existing checkpoint (#3824 by @AHSharan).
Outlook: PEFT v1.0
The next PEFT release will be the 1.0 release! For this, we have planned a couple of sweet additions that will extend the scope of what PEFT is currently capable of doing, so stay tuned.
We will also take the opportunity to clean up PEFT and to streamline the API. This will mostly concern PEFT internals and we will take great care of ensuring backwards compatibility in the places where it matters, as well as to keep existing checkpoints working. Still, expect a few smaller breaking changes for the more uncommon workflows. Just as an example: Right now, if you have a PEFT model with multiple active adapters, and then you merge a single of those adapters, all other active adapters are completely ignored. This is very surprising and will raise an error in the future. For most PEFT users, however, everything will keep working as expected. We will have a release candidate so that you can check if anything breaks and give us feedback.
All changes
- CHORE Update triage bot for GitHub PRs by @BenjaminBossan in #3749
- Key the doc build concurrency group on the pull request number by @albertvillanova in #3713
- LoRA TP DTensor migration by @michaelbenayoun in #3614
- Benchmark results 0.21.0 by @githubnemo in #3754
- fix(delete): restore fallback trainability after deleting active adapter by @Sravanjangam in #3743
- Bump dev version by @githubnemo in #3750
- fix: store miss_fn per-adapter in MiSS layer by @peft-jambot in #3379
- Small helper to run experiments on HF jobs by @githubnemo in #3666
- FIX add_weighted_adapter creates an overscaled adapter when the source uses rslora by @sohumt123 in #3449
- Fix rank and alpha pattern matching during LoRA hotswap by @DaoyuanLi2816 in #3752
- FIX KappaTune does not dequantize bitsandbytes 8-bit weights by @amogh-nagri-11 in #3759
- CI Skip workflows that need repo secrets when running on forks by @Ayushraj06-bit in #3732
- Add VeLoRA experiment (LoRA r=32 + velora_config) to image-gen benchmark by @Sergasgr in #3764
- Avoid duplicated shared projections in RandLoRA adapter checkpoints by @CoralGarden52 in #3714
- fix: stop task-type PeftModels from mutating the caller's modules_to_save by @Vedant-Agarwal in #3770
- FIX get_delta_weight for pointwise Conv3d layers in LoRA and HiRA by @Nicholas022400701 in #3777
- CHORE Update AGENTS.md and test instructions by @BenjaminBossan in #3782
- fix(dora): use base layer device instead of hardcoding cuda in ephemeral offload by @li-lizhe in #3734
- DOC Merging a subset of the active adapters by @AHSharan in #3763
- FIX Preserve Conv LoRA dilation and padding_mode by @TomasGuija in #3758
- Proper handling of named device mesh for TP by @michaelbenayoun in #3790
- Add BD-LoRA experiment to image-gen benchmark by @Sergasgr in #3799
- Add MonteCLoRA experiment to image-gen benchmark by @Sergasgr in #3806
- TST Add multihead attention model to regression tests by @BenjaminBossan in #3778
- CHORE Triage bot: enable cron, issue allow list by @BenjaminBossan in #3781
- TST Improve test coverage prompt learning by @BenjaminBossan in #3570
- CI Remove mslk from nightly CI by @BenjaminBossan in #3810
- CHORE Triage: exempt hf security analysis bot by @BenjaminBossan in #3812
- Fix Trainable Tokens output-head bias by @DaoyuanLi2816 in #3688
- FIX Match hotswap adapters by exact name by @aswanth-07 in #3794
- docs: document Riemannian-preconditioned LoRA optimizer by @smellslikeml in #3751
- FEAT Add Astra initialization for LoRA by @LyoAI in #3725
- fix: drop obsolete encoder_decoder kwargs patch in seq2seq models by @peft-jambot in #3845
- FIX modules_to_save KeyError on params-only state_dict by @rupeshpoojary9 in #3816
- Add TinyLoRA experiment to image-gen benchmark by @Sergasgr in #3850
- method_comparison/image-gen: set a usable LR for ln_tuning by @zdd3143 in #3849
- Fix nested Trainable Tokens wrappers on tied output heads by @DaoyuanLi2816 in #3820
- Bump version to 0.21.3.dev0 by @BenjaminBossan in #3856
- Add Super-Tuning experiment to image-gen benchmark by @Sergasgr in #3852
- CHORE Make triage script more robust by @BenjaminBossan in #3851
- FIX Adapters stay frozen after disable_adapter() with FSDP2 by @hivaze in #3839
- FIX
modules_to_savebuffer handling inconvert_to_loraby @eSVeeF in #3481 - method_comparison: implement pattern matching in search by @githubnemo in #3893
- Fix Trainable Tokens adapters with disjoint target modules by @DaoyuanLi2816 in #3848
- FIX Restore inference mode when saving adapter config fails by @DaoyuanLi2816 in #3866
- FIX Pattern key warning ignores target_parameters by @surajsharan in #3891
- CI Fix several nightly GPU run errors by @BenjaminBossan in #3870
- FEAT Add a 'healthcheck' method to PEFT models by @BenjaminBossan in #3798
- CI: Bump the third-party-actions group with 6 updates by @dependabot[bot] in #3853
- FIX LoHa wrong gradients with use_effective_conv2d=True by @Sergasgr in #3869
- DOC Note that unmerging quantized weights is not exact by @chrikrah in #3894
- FIX modules_to_save must match whole module names by @AHSharan in #3824
- CI HF doc-builder now needs read permission by @BenjaminBossan in #3909
- FIX MultiheadAttention LoRA merge() applies out_proj delta twice by @Vedant-Agarwal in #3774
- Add VeLoRA experiment to MetaMathQA benchmark by @Sergasgr in #3912
- ENH Optimize HRA forward for better memory efficiency by @BenjaminBossan in #3185
- CI HF pr-doc-builder now needs read permission by @BenjaminBossan in #3916
- Fix seq2seq prompt generation method restoration by @DaoyuanLi2816 in #3897
- FIX MoE weight conversion issue with ambiguous name by @BenjaminBossan in #3715
- Fix prompt learning generate warning about position_ids at every decoding step by @albertvillanova in #3917
- FIX set_adapter leaves the new adapter frozen after a merge under FSDP2 by @rupeshpoojary9 in #3895
- Update FourierFT image-gen experiment with tuned learning rate by @Sergasgr in #3933
- Update LoHa image-gen experiment with tuned learning rate by @Sergasgr in #3934
- Method comparison: Fix run.py crash after the healthcheck change by @Sergasgr in #3936
- Fix OSF merge applying the delta twice on repeated merge by @adi-IL in #3886
- FIX format_healthcheck ignores sep and indent by @hpdkhoa in #3940
- FIX add_weighted_adapter SVD combination types for Conv1d and Conv3d LoRA layers by @Nicholas022400701 in #3903
- Add Relation Extraction LoRA script and notebook tutorial by @GauravPatil2515 in #3368
- FIX Keep the order of rank_pattern and alpha_pattern keys on save by @jayzuccarelli in #3947
- CHORE Python 3.10 is EOL, remove it, add 3.14 by @BenjaminBossan in #3939
- CHORE Triage bot: Make it possible to assign contributor by name by @BenjaminBossan in #3971
- FIX Save/load with multiple adapters when only one uses trainable tokens by @superdiaodiao in #3950
- FIX Clarify PeftModel.set_adapter errors for non-string names by @matheusht in #3902
- Keep dtype and device when loading an adapter with low_cpu_mem_usage by @vineethsaivs in #3509
- CHORE Add Pauline to allowed users for triage bot by @BenjaminBossan in #3904
- FIX OFT and GraLoRA with Linear inputs of more than 3 dims by @mottopanikeiku in #3974
- FIX MoE conversion skipping routed experts when shared experts or dense layers share their name by @albertvillanova in #3946
- Add supertuning tests for other platform like XPU as well by @kaixuanliu in #3975
- FIX Respect adapter module training state in explicit inference mode by @eSVeeF in #3809
- FIX enable gradients in LoRA-FA optimizer closures by @cjwystz in #3966
- FIX CPT forward mutates caller-provided input_type_mask in place by @n-dlms in #3969
- FIX LN Tuning KeyError when the active adapter does not target a layer by @behroozazarkhalili in #3980
- FIX Old, unused torchao install breaking adapter injection by @Vedavarshith868 in #3989
- FIX
unload()returns the trainedmodules_to_savecopy instead of the original module by @Sergasgr in #3957 - Filter adapter-name slots by tuner-prefix position by @Sravanjangam in #3623
- FIX Freeze inference-only prompt adapter parameters by @DaoyuanLi2816 in #3907
- Fix LoRA-FA updates for adapters without gradients by @DaoyuanLi2816 in #3972
- Update LoHa MetaMathQA experiment with tuned learning rate by @Sergasgr in #4005
- Only gather adapter weights when saving a DTensor TP model by @michaelbenayoun in #3963
- docs: Add basic working principle for CPT by @githubnemo in #3978
- AI: Further extend AI agent instructions by @BenjaminBossan in #4009
- Pin UTF-8 encoding when reading/writing adapter config JSON by @dltsum in #3771
- FIX unload crash for composite modules_to_save without weight by @Sergasgr in #3937
- fix(bnb): apply scaling to lora_bias merge/unmerge on 8-bit and 4-bit by @Sravanjangam in #3757
- FIX Raise an error when a PEFT method targets an unsupported layer type by @devYRPauli in #4015
- FIX add_weighted_adapter for lora_bias, DoRA and KaSA by @LK-maker-007 in #3844
- CHORE Triage: fix wrongly documented command by @BenjaminBossan in #4021
- FIX Raise an error for lora_bias=True on MultiheadAttention by @Vedant-Agarwal in #4001
- [DOC] Fix the minimum version of Python required by @NoeFlandre in #4014
- FIX all-linear skipping dense and shared-expert MoE layers by @rupeshpoojary9 in #3819
- DOC Simplify optimized-path troubleshooting wording by @gorkadabo1 in #4025
- FIX safe_merge leaving the base layer modified, and merging several adapters at once on torchao by @yupengtang in #3779
- FIX LoRA initialization of Conv1D layers for OLoRA, LoftQ, LoRA-GA and BD-LoRA by @mottopanikeiku in #4012
- FIX Use the active CPT adapter config for loss by @DaoyuanLi2816 in #3941
- FIX make adapter injection atomic on validation failure by @eSVeeF in #3981
- FIX Preserve trainability when loading mixed adapters by @aswanth-07 in #3862
- NoRA-init by @Joluck in #3908
- FIX LoftQ init failing for Conv, Embedding, target_parameters and Megatron layers by @nettanvirdev in #4029
- FEAT Add PaCA (Partial Connection Adaptation, ICLR 2025) by @WooSunghyeon in #3927
- Use the peft-internal-testing tiny random Llama model by @albertvillanova in #4056
- CI: Make HF_HOME an absolute path by @githubnemo in #4060
- TST: avoid Hub requests when custom model tests save adapters by @peft-jambot in #4061
- TST Fix flaky Hub-related test failures by @rtrompier in #4058
- FIX Avoid making ranknum in AdaLoRA trainable by accident by @BenjaminBossan in #3859
- TST Add tests for merging LoRA with tied weights by @BenjaminBossan in #3954
- FIX TST Flawed BEFT checking logic in test by @BenjaminBossan in #4038
- FIX Reshard on disable_adapter() exit only when needed by @hivaze in #3938
- FIX Adapter shape for FSDP2 layers on meshes not named "fsdp" by @hivaze in #3838
- FIX Outdated imports and config args in AdaMSS quick test script by @sarthakuwar in #4073
- FIX Several fixes to Conv2d layers, extend test coverage by @BenjaminBossan in #3837
- Bump doc-builder to get the docs machine-translated by @mishig25 in #4072
- Broadcast full adapter weights by @michaelbenayoun in #3965
- FIX OFT merge order for multiple active adapters by @anh-nguyenn in #4068
- CI: Bump actions/setup-python from 7.0.0 to 7.1.0 in the ci-actions group by @dependabot[bot] in #4092
- CI Add nightly CI workflow that runs on MacOS with Apple Silicon by @BenjaminBossan in #4046
- Fix broken examples on transformers 5.x (target_modules, dataset IDs, warmup_ratio) by @apuravmanhas in #4086
- Fix modules_to_save selection when merging an inactive adapter by @gorkadabo1 in #4085
- FEAT healthcheck: warn when a randomly initialized task head is not saved by @eSVeeF in #4089
- FIX reduce_intruder_dimension for fan_in_fan_out and embedding LoRA layers by @vineethsaivs in #4100
- FIX macOS test failures and MPS float64 errors in FRoD and EVA by @Alex-Wengg in #4102
- Chore transformers renames gradient checkpointing to activation checkpointing by @BenjaminBossan in #4114
- [Fixes #4101] Support
torch.acceleratorbackends (e.g.torch_tpu) ininfer_device()by @jigyasa-grover in #4099 - DOC Fix docstring examples of LoHa, LoKr, OFT, MiSS and HRA by @blipbyte in #4091
- FIX Validate the requested adapter configuration during hotswap by @DaoyuanLi2816 in #3865
- Use a denylist for Conv1D test skips by @Sergasgr in #4075
- docs: add use cases and limitations to Llama-Adapter page by @sarvadnya2030 in #4103
- FIX Reject layers_to_transform with a regex str target_modules for all PEFT methods by @Shaurya55555 in #4087
- Use the peft-internal-testing tiny Llama4 model by @albertvillanova in #4057
- Use the peft-internal-testing tiny random Llama model in the parallelism tests by @albertvillanova in #4121
- FIX Gather FSDP2 adapter DTensors when saving by @Levius-Fubuki in #4065
- FIX AdaLoRA Conv2d with groups and padding_mode by @Sergasgr in #4084
- FIX strict_adapter_check for load_adapter and from_pretrained by @rupeshpoojary9 in #3832
- CI: Better slow test report by @BenjaminBossan in #4136
- ENH Add replace_lora_weights_pissa for 4-bit models by @jietinatian0803 in #4135
- TST Replace tiny TRL models with PEFT models by @BenjaminBossan in #4141
- docs: update README quickstart text to match Qwen example by @AnandaRimal in #4145
- ENH: Svd lowrank option for lora intruder dimension reduction by @causeri3 in #4081
- [Distributed]
embedding_colwisefor LoRA adapters by @michaelbenayoun in #4112 - FIX Empty adapter state dict with nested FSDP or checkpoint_wrapper by @hivaze in #4108
- save_and_load: fix false-positive warning with LoRA bias by @githubnemo in #3817
- docs: assorted fixes by @githubnemo in #4124
- TST Speed up three tests that are slow on aggregate by @BenjaminBossan in #4137
- FIX all-linear skips experts on qwen2_moe after transformers v5 conversion by @hivaze in #4154
- ENH Scale delta weights in place when merging by @Hang-get in #4147
- AGENTS.md: PR description should fit git history by @githubnemo in #4034
- FIX prepare_model_for_kbit_training defaults to use_reentrant False by @BenjaminBossan in #4146
- feat: support LoRA conversion for multiplicative PEFT methods (OFT, BOFT) by @peft-jambot in #3360
- FEAT Add helper function to create a dummy LoRA config to hotswap by @BenjaminBossan in #3739
- Bump dev version to 0.21.4.dev by @githubnemo in #4167
- FIX Make FSDP2 merges reach the sharded parameters by @RerankerGuo in #4077
- FIX BOFT and OFT Conv2d layers by @MaiAditya in #4157
- CHORE Stale bot exempts post-v1 label by @BenjaminBossan in #4179
- Update FourierFT MetaMathQA experiment with tuned learning by @Sergasgr in #4182
- Update LN Tuning MetaMathQA experiment with tuned learning rate by @Sergasgr in #4183
- DOC Update FSDP guide and add an FSDP2 SFT configuration by @DaoyuanLi2816 in #4107
- FIX Raise helpful error when merging prompt learning or adaption prompt methods by @RerankerGuo in #4180
- FIX Preserve CPT input token roles by @aswanth-07 in #3953
- Fix HRA Conv2d string padding by @gorkadabo1 in #4159
- Fix PeftMixedModel.disable_adapter state restoration by @YueshenZ in #3910
- Re-enable Transformer Engine in GPU Docker image by @rahulcomet in #4186
- FIX Failing OFT and BOFT conv padding tests by @BenjaminBossan in #4188
- FIX Healthcheck now also works for prompt learning by @BenjaminBossan in #3944
- FIX MHA weight restoration issues by @BenjaminBossan in #4083
- FIX Broken MHA FSDP test by @BenjaminBossan in #4190
- FIX Transformers deprecates past_key_values in favor of cache by @BenjaminBossan in #4165
- docs: add FourierFT selection guidance by @gurinderpreetbrraich-cyber in #4187
New Contributors
- @sohumt123 made their first contribution in #3449
- @Ayushraj06-bit made their first contribution in #3732
- @Nicholas022400701 made their first contribution in #3777
- @li-lizhe made their first contribution in #3734
- @AHSharan made their first contribution in #3763
- @TomasGuija made their first contribution in #3758
- @LyoAI made their first contribution in #3725
- @rupeshpoojary9 made their first contribution in #3816
- @zdd3143 made their first contribution in #3849
- @hivaze made their first contribution in #3839
- @surajsharan made their first contribution in #3891
- @chrikrah made their first contribution in #3894
- @adi-IL made their first contribution in #3886
- @hpdkhoa made their first contribution in #3940
- @GauravPatil2515 made their first contribution in #3368
- @jayzuccarelli made their first contribution in #3947
- @superdiaodiao made their first contribution in #3950
- @matheusht made their first contribution in #3902
- @vineethsaivs made their first contribution in #3509
- @mottopanikeiku made their first contribution in #3974
- @cjwystz made their first contribution in #3966
- @n-dlms made their first contribution in #3969
- @Vedavarshith868 made their first contribution in #3989
- @dltsum made their first contribution in #3771
- @devYRPauli made their first contribution in #4015
- @NoeFlandre made their first contribution in #4014
- @gorkadabo1 made their first contribution in #4025
- @yupengtang made their first contribution in #3779
- @nettanvirdev made their first contribution in #4029
- @WooSunghyeon made their first contribution in #3927
- @sarthakuwar made their first contribution in #4073
- @anh-nguyenn made their first contribution in #4068
- @apuravmanhas made their first contribution in #4086
- @Alex-Wengg made their first contribution in #4102
- @jigyasa-grover made their first contribution in #4099
- @sarvadnya2030 made their first contribution in #4103
- @Shaurya55555 made their first contribution in #4087
- @Levius-Fubuki made their first contribution in #4065
- @jietinatian0803 made their first contribution in #4135
- @AnandaRimal made their first contribution in #4145
- @causeri3 made their first contribution in #4081
- @Hang-get made their first contribution in #4147
- @RerankerGuo made their first contribution in #4077
- @MaiAditya made their first contribution in #4157
- @YueshenZ made their first contribution in #3910
- @rahulcomet made their first contribution in #4186
- @gurinderpreetbrraich-cyber made their first contribution in #4187
Full Changelog: v0.21.3...v0.22.0