unstale

pypi / timm

timm

What changed in timm 1, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-24.

2 breaking1 security19 new APIs and options3 performance1 behavior and defaults31 fixes and other

See which of these apply to your code, ranked, in your agent:

npx @unstale/cli init

@unstale/cli 0.2.4 on npm

timm 157 changes

Breaking changes 2

  • Changes input-size selection and fixes test-time preprocessing defaults.

    1.0.30release notes

  • Switches the default position interpolation.

    1.0.30release notes

Security 1

  • Pickle loading is hardened to address security concerns.

    1.0.28release notes

New APIs and options 19

  • Adds DeepSeek-V4 and V4.1 vision classifiers and encoders with native timm Hub weights.

    1.0.30release notes

  • Adds EfficientViM with native timm Hub weights.

    1.0.30release notes

  • Adds iFormer with native timm Hub weights.

    1.0.30release notes

  • Adds Qwen-Drive-1.0-4B vision tower weights as a qwen3_vit pretrained tag.

    1.0.30release notes

  • Adds Qwen3.5 and Qwen3.8 vision encoders, and the Qwen-Drive-1.0-4B vision tower pretrained tag.

    1.0.30release notes

  • Adds ViT classifier and encoder variants, including classifiers with and without the native spatial merger.

    1.0.30release notes

  • Adds ViT classifier and encoder variants, including classifiers with and without the native spatial merger.

    1.0.30release notes

  • Adds ViT classifier and encoder variants, including classifiers with and without the native spatial merger.

    1.0.30release notes

  • Adds ViTs with EVA and NaFlexViT support and native timm Hub weights.

    1.0.30release notes

  • Adds LingBot-Vision models and pretrained weights.

    1.0.29release notes

  • Adds optional key-only masks for NaFlexVit self-attention.

    1.0.29release notes

  • Adds per-batch image and batch size scheduling for non-NaFlex training, including progressive small-to-large resolution schedules.

    1.0.29release notes

  • Adds PP-LCNetV2 model definitions and pretrained weights, including small, base, and large variants.

    1.0.29release notes

  • Model definitions and pretrained weights are added for EUPE ViT and ConvNeXt models.

    1.0.28release notes

  • Model definitions and pretrained weights are added for EUPE ViT and ConvNeXt models.

    1.0.28release notes

  • NaFlexViT patch-layout support is extended, including support for NaFlex-CLAP.

    1.0.28release notes

  • Adds a flag to unify channels-last patchification support across the data pipeline and NaFlex configuration and models.

    1.0.28release notes

  • TIPSv2 model definitions and pretrained weights are added for DINOv2-style ViTs.

    1.0.28release notes

  • Adds a seed option for reproducible validation.

    1.0.28release notes

Performance 3

  • Adds caching for variable-patch interpolation.

    1.0.29release notes

  • Makes position embedding sampling compile-friendly and memory-efficient.

    1.0.29release notes

  • Returns the fused RMSNorm2d result directly.

    1.0.29release notes

Behavior and default changes 1

Fixes and other changes 31

  • Fixes RMS clipping direction.

    1.0.30release notes

  • Fixes learning-rate scaling for convolution weights.

    1.0.30release notes

  • Preserves the caller's gradient mode.

    1.0.30release notes

  • Clears the eval-mode cache on every train() or eval() call to prevent stale attention bias caches.

    1.0.30release notes

  • Fixes mixing buffer shape for non-square and single-band images.

    1.0.30release notes

  • Adds device/dtype factory support for a few models.

    1.0.30release notes

  • Fixes channel metadata for pruned models.

    1.0.30release notes

  • Corrects weight decay when loading checkpoints.

    1.0.30release notes

  • Copies the previous gradient instead of retaining a reference to p.grad.

    1.0.30release notes

  • Fixes models that cannot be created with a meta device context.

    1.0.30release notes

  • Improves parsing, including values containing '=' or non-literals.

    1.0.30release notes

  • Fixes stream reuse and synchronization.

    1.0.30release notes

  • Fixes padding for dilated max pooling.

    1.0.30release notes

  • Fixes factored row normalization for tall matrices.

    1.0.29release notes

  • The b config flag now honors the documented value 0.

    1.0.29release notes

  • Fixes an AttributeError when hparams is None.

    1.0.29release notes

  • The inc config flag now honors the documented value 0.

    1.0.29release notes

  • Classifier resets preserve the head device and dtype.

    1.0.29release notes

  • Fixes device and dtype changes for all classifier resets.

    1.0.29release notes

  • Fixes minmax border sampling and additional mixup, CutMix, and bounding-box range issues.

    1.0.29release notes

  • Fixes the NaFlexVit load_pretrained helper and clarifies its purpose.

    1.0.29release notes

  • Improves model factory path and checkpoint handling.

    1.0.29release notes

  • Fixes a crash when layer_decay is set without an explicit minimum scale.

    1.0.29release notes

  • Fixes positional embedding resampling when grid changes have an equal number of tokens.

    1.0.29release notes

  • Fixes handling of custom input channels for pruned EfficientNet models.

    1.0.29release notes

  • Fixes the foreach path so weight decay is not applied once per device and dtype group.

    1.0.29release notes

  • Fixes ZeroDivisionError when the count is 0.

    1.0.28release notes

  • Inference labels are fixed for models fine-tuned on non-ImageNet datasets.

    1.0.28release notes

  • Fixes handling so reg_token is included in no-weight-decay and stem/embed learning-rate groups.

    1.0.28release notes

  • Support for tensor learning rates is improved.

    1.0.28release notes

  • Optimizer support for torch.compile is improved by using a tensor for the step count.

    1.0.28release notes