build(deps): bump tabicl from 2.1.1 to 2.2.0

[dependabot]

Bumps tabicl from 2.1.1 to 2.2.0.

Release notes

Sourced from tabicl's releases.

v2.2.0

What's Changed

• add pytest to "test" dependency group in pyproject.toml by @​jeromedockes in soda-inria/tabicl#122
• only set model_ when unpickling if pickled model was fitted by @​jeromedockes in soda-inria/tabicl#121
• fix get/set state model_kv_cache_ by @​jeromedockes in soda-inria/tabicl#124
• fix default behaviour on NumPy array with strings by @​marineLM in soda-inria/tabicl#123
• n_threads is set to the maximum of n_logical_cores and n_jobs, rather than minimum by @​douglas-boubert in soda-inria/tabicl#107
• Prior + pre-training code by @​dholzmueller in soda-inria/tabicl#135
• Document skrub.tabular_pipeline's new support for TabICL by @​ashwinvis in soda-inria/tabicl#139
• Fix UnicodeEncodeError in fine-tuning tutorials on Windows by @​maxdemarzi in soda-inria/tabicl#141
• Fix broken star history chart in README by @​FaintFlower in soda-inria/tabicl#149
• Keep all-NaN columns in SimpleImputer so predict_proba does not crash by @​axsaucedo in soda-inria/tabicl#148
• doc(readme): remove ambiguity on classification/regression by @​MarieSacksick in soda-inria/tabicl#134
• fix: use mix_probs key in SCMPrior by @​nightcityblade in soda-inria/tabicl#106
• [EHN] Remove gluonts dependency by @​daidahao in soda-inria/tabicl#108
• Change config variables names in prior by @​ChristopheMuller in soda-inria/tabicl#105
• Refactor device handling to make it backend agnostic by @​ogrisel in soda-inria/tabicl#144
• Keep one constant feature so all-constant X fits instead of raising by @​klemens-floege in soda-inria/tabicl#150
• Multiple improvements / bug fixes by @​dholzmueller in soda-inria/tabicl#151
• Use grad scaler for float32 + FlashAttention-3 by @​dholzmueller in soda-inria/tabicl#154
• Respect cgroup memory limits when detecting available CPU memory by @​ogrisel in soda-inria/tabicl#152

New Contributors

• @​jeromedockes made their first contribution in soda-inria/tabicl#122
• @​douglas-boubert made their first contribution in soda-inria/tabicl#107
• @​ashwinvis made their first contribution in soda-inria/tabicl#139
• @​maxdemarzi made their first contribution in soda-inria/tabicl#141
• @​FaintFlower made their first contribution in soda-inria/tabicl#149
• @​axsaucedo made their first contribution in soda-inria/tabicl#148
• @​MarieSacksick made their first contribution in soda-inria/tabicl#134
• @​nightcityblade made their first contribution in soda-inria/tabicl#106
• @​daidahao made their first contribution in soda-inria/tabicl#108
• @​ChristopheMuller made their first contribution in soda-inria/tabicl#105
• @​klemens-floege made their first contribution in soda-inria/tabicl#150

Full Changelog: soda-inria/tabicl@v2.1.1...v2.2.0

Changelog

Sourced from tabicl's changelog.

2.2.0

New features

• Add TabICLv2 pre-training code (python -m tabicl.train): quantile regression training via a pinball loss (--regression_method quantile) in addition to classification, and the Muon optimizer (--muon True) alongside AdamW. The training CLI now also exposes the graph_scm prior options, layernorm-without-bias (--norm_type layernorm_nobias), SSMax (--col_ssmax/--icl_ssmax with --ssmax_type), feature grouping and target-aware embeddings (--col_feature_group, --col_target_aware, --col_affine), the RoPE variant (--row_rope_interleaved; v1 interleaved by default, v2 uses False), residual initialization (--zero_init; v2 uses False), FlashAttention-3 during training (--use_flash_attn3; the v2 recipe enables it for stages 2 and 3 only). The cuDNN SDPA backend is now automatically disabled during training (slower than Flash Attention on Hopper). --dtype now supports bfloat16 in addition to float16/float32; the GradScaler is only enabled when the computation graph contains a float16 path (either float16, or float32 with FlashAttention-3 enabled since it runs in float16 then). All CLI defaults reproduce the TabICLv1 model configuration; resuming a run re-seeds the data stream with the current step. (PR#135)

• Remove the GluonTS dependency from the forecasting module. (PR#108, @​daidahao)

• Improve non-CUDA GPU inference reliability and performance (including XPU and Apple Silicon MPS): inference now consistently runs on the configured backend device, uses backend-appropriate autocast, and queries available memory plus async stream/event primitives through backend-agnostic torch.<backend> APIs (with safe synchronous fallbacks when async is unavailable). This fixes pathological auto-batch sizing (e.g. batch size forced to 1) and restores expected accelerated inference behavior on supported non-CUDA GPU backends. MPS now uses the same AMP, auto-batching, and memory-aware inference path as other accelerators instead of falling back to the CPU path. use_amp="auto" is device-aware (off on CPU; size heuristic on CUDA/XPU/MPS), and float16 KV caches are kept on MPS when AMP is enabled. When device=None, estimators now default to CUDA when available, otherwise XPU, then MPS, and then CPU. (PR#144)

Bug fixes

• Respect cgroup memory limits when estimating available CPU memory at inference, so offload="auto" inside Docker or Kubernetes does not treat host RAM as free and get OOM-killed with no traceback. (PR#152)

• Fix the pre-training GradScaler being disabled for --dtype float32 --use_flash_attn3 True (the TabICLv2 stage 2 and 3 recipes), which degraded pre-training quality: FlashAttention-3 runs its backward pass in float16 there, so loss scaling is still needed. (reported by @​wangzhengli in #153)

• Finetuning now supports string/categorical features in DataFrames, matching the behavior of the base TabICLClassifier and TabICLRegressor. Previously, FinetunedTabICLClassifier and FinetunedTabICLRegressor would raise ValueError: could not convert string to float when the input contained categorical columns. (PR#151; reported by @​zehua-jerry-yu in #118)

• Fix DatetimeEncoder sin/cos encoding off-by-one error that caused the first and last elements of a period to map to identical angles (e.g. Monday and Sunday getting the same encoding). The denominator was incorrectly p-1 instead of p. (PR#151; reported by @​christophM in #136)

• Fix float16 input arrays crashing during Yeo-Johnson normalization in the preprocessing pipeline. The PreprocessingPipeline now upcasts float16 to float32 before fitting/transforming, avoiding scipy's narrow-exponent bound error. (PR#151; reported by @​SebastienMelo in #140)

• Fix predict_proba/predict crashing with TypeError when a categorical column is all-NaN in the prediction batch. Removed the batch-global all-NaN feature-masking detection from the prediction path — it was intended for SHAP but did not work correctly with SHAP's coalition batching, and made predictions depend on batch composition. All-NaN columns now flow through normal preprocessing (OrdinalEncoder/SimpleImputer handle NaN natively). (PR#151; reported by @​Innixma in #143)

• Fix PyTorch autograd error when fine-tuning with partial module freezing (e.g. freeze_col=True, freeze_row=True, freeze_icl=False). An in-place operation on a tensor view from frozen modules conflicted with autograd; resolved by detaching before the in-place write. (PR#151; reported by @​denisfouchard in #128)

• When unpickling a TabICL estimator, the fitted attributes device_, model_, etc. are only set if the pickled model was fitted. (PR#121, @​jeromedockes)

• Fix get_state/set_state for model_kv_cache_. (PR#124, @​jeromedockes)

... (truncated)

Commits

• 0dbff3e Respect cgroup memory limits when detecting available CPU memory (#152)
• 3c82053 Use grad scaler for float32 + FlashAttention-3 (#154)
• de37591 Multiple improvements / bug fixes (#151)
• 8ac8ca8 Keep one constant feature so all-constant X fits instead of raising (#150)
• 8f1aa20 Refactor device handling to make it backend agnostic (#144)
• 7aae2dc Change config variables names in prior (#105)
• 7dea708 [EHN] Remove gluonts dependency (#108)
• bc769de fix: use mix_probs key in SCMPrior (#106)
• 90e3a2b doc(readme): remove ambiguity on classification/regression (#134)
• cb92a63 Keep all-NaN columns in SimpleImputer so predict_proba does not crash (#148)
• Additional commits viewable in compare view

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

---

Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

• @dependabot rebase will rebase this PR
• @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
• @dependabot show <dependency name> ignore conditions will show all of the ignore conditions of the specified dependency
• @dependabot ignore this major version will close this PR and stop Dependabot creating any more for this major version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this minor version will close this PR and stop Dependabot creating any more for this minor version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this dependency will close this PR and stop Dependabot creating any more for this dependency (unless you reopen the PR or upgrade to it yourself)

[amrit110]

Automated fix applied and PR merged

The agentic fix loop successfully fixed this PR and merged it.

✓ Successfully fixed merge_only failures - Modified 0 files - Executed 102 agent actions - (55 info, 22 tool_call, 3 error, 18 tool_result, 4 reasoning)

View PR history on dashboard

AI Engineering Maintenance Bot