$ the-wire · showcase
kernels stops reloading duplicate kernels, restores pre-2.10 Torch CPU/ROCm builds
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
huggingface/kernels fixed redundant kernel loading that caused naming clashes and restored CPU/ROCm builds for Torch versions older than the 2.10 stable ABI, while the day's remaining commits stayed inside CI and generated docs.
kernels: do not reload kernel with the same ID huggingface/kernels
When a kernel with a given ID was already loaded, the runtime used to load it again; it now skips the reload. The maintainers note the repeat load can be costly and can cause naming clashes, so callers that loaded the same kernel twice will see fewer duplicate loads in process.
Add back Torch 2.10-2.12 for CPU/ROCm huggingface/kernels
Older CPU/ROCm Torch versions were added back for the case where a build cannot target the Torch 2.10 stable ABI. The corresponding commit carries the same restore, covering kernels that need to run against Torch 2.10 through 2.12 on CPU and ROCm.
Fix test selection for virtual module imports (#49690) huggingface/transformers
The tests_fetcher.py dependency map used to treat transformers.tokenization_utils_fast as a file dependency, though that legacy module is served dynamically, which raised a ValueError and derailed test selection; the mapping now handles the dynamic import. This is CI plumbing, not a runtime change to the library.
[Bot] Update Inference Providers documentation huggingface/hub-docs
The bot workflow bumped @huggingface/tasks and @huggingface/inference to latest and regenerated the Inference Providers docs, and the change landed as a matching commit. Nothing to read unless you track the published provider list.
