$ the-wire · showcase
Transformers drops legacy data helpers, TRL rebuilds GOLD tests on real trainers
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
A housekeeping day across the Hugging Face stack: Transformers removed years-deprecated GLUE/SQuAD/XNLI data helpers and stray cross-model imports, while TRL replaced stub-built GOLD trainers with the real thing and kernels-community landed batched-generation fixes.
Remove deprecated data helpers (#49360) huggingface/transformers
The pre-`datasets` helpers (`GlueDataset`, `SquadDataset`, the GLUE/SQuAD/XNLI processors and metrics) are gone from `transformers.data.{datasets,processors,metrics}`, along with their top-level exports and dummy objects; the data collators and the slow GLUE evaluation set are unchanged. Marc Sun notes the helpers "have been deprecated for years and have had no users in the library since the le...
Build real GOLD trainers in the on-policy generation tests huggingface/trl
The GOLD on-policy generation tests now run against a real trainer with its real tokenizer and model instead of a `__new__`-built trainer, hand-written tokenizers and a scripted model. The stubs could not notice a change in trainer setup or in how a real tokenizer decodes special and pad tokens, so regressions in that path are now visible in CI.
Build real GOLD VLM trainers in the on-policy generation tests huggingface/trl
The same treatment for the GOLD vision-language generation tests, which previously skipped `__init__`, generated with a fake model, used placeholder image objects and patched `unwrap_model_for_generation` and `prepare_multimodal_messages`. Third in the set, alongside the input-preparation tests and the contributing-guide guidance on testing through the public API, all under the #7565 effort.
weathernext2-banded-attention: follow torch's float32 matmul precision (TF32x3 on NVIDIA) huggingface/kernels-community
The model-facing attention in weathernext2-banded-attention now follows PyTorch's float32 matmul setting rather than always using IEEE dots: `tf32x3` at default full precision on NVIDIA only, Triton's default TF32 when TF32 is allowed, and `ieee` on ROCm and XPU. The change was requested by the kernel's author in the corresponding Transformers thread.
Update mlx-quantization-metal-kernels huggingface/kernels-community
Also in kernels-community: mlx-quantization-metal-kernels was synced with the latest MLX quantization ops, fixing errors in batched generation, and ggml-gated-delta-net fixed `delta_gates` for batched decode. In huggingface/kernels, the nix-builder added Torch 2.15.0rc3 support along with new Triton dependencies and cupti-python for Torch 2.15 on CUDA 13.4.
Action items
- → Replace any imports of `transformers.data.datasets`, `transformers.data.processors` or `transformers.data.metrics` (G... huggingface/transformers [plan]
