8 pull requests merged across the ML ecosystem.
NVIDIA-NeMo/Automodel · NVIDIA: Fix Gemma4 SDPA NaN by running attention in fp32 on Hopper
rapidsai/cuml · NVIDIA RAPIDS: Add regression test for cluster-metric label-permutation invariance
scikit-learn/scikit-learn: Support mixed namespace/device Array API inputs in PoissonRegressor
sktime/sktime: Add HyperTreeNetAR forecaster interface to hypertrees-forecasting
Farama-Foundation/Gymnasium: Use trained-policy GIFs for Box2D environments (LunarLander, BipedalWalker)- Farama-Foundation/Gymnasium: Use trained-policy GIFs for classic control environments
mlflow/mlflow: Fix promo modal focus (aria-hidden) and show it on the empty models page
Observal/Observal: Add a harness support GitHub issue form