$ the-wire · showcase
SageMaker S3 ownership guard, kernels breaks setuptools builds
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
11 people shipped this
Xinhao Wei
@XiaomingFun233
1 cited
Cyril Vallez
@Cyrilvallez
1 cited
openvino-agent
@openvino-agent
1 cited
anatyrova
@anatyrova
1 cited
rsareddy0329
@rsareddy0329
1 cited
papriwal
@papriwal
1 cited
dwarez
@dwarez
1 cited
jiqing-feng
@jiqing-feng
1 cited
SunMarc
@SunMarc
stevhliu
@stevhliu
molbap
@molbap
Two fixes land today with consequences outside their own repos: the SageMaker Python SDK now verifies it owns the S3 bucket it writes to, and a bad entry point in kernels 0.17.0 is fixed after breaking every setuptools source build in an affected environment.
fix(train): enforce S3 ownership on ai_registry default bucket huggingface/sagemaker-python-sdk
The ai_registry and finetune_utils modules derived a predictable default bucket name, sagemaker-{region}-{account_id}, and read and wrote to it without checking ownership, so another account could pre-create that globally unique name and the SDK would silently use it. The guard now applies to the SDK-derived default bucket; explicitly provided buckets are left untouched.
Fix `egg_info.writers` entry point module path (#829) huggingface/kernels
A wrong egg_info.writers entry point path meant that with kernels 0.17.0 installed, any setuptools source build in the same environment failed with ModuleNotFoundError: No module named 'kernels.lockfile', including builds of unrelated packages. Pinning or upgrading matters if you build from source in an environment where kernels is present.
[generate] Drop attention mask early without padding (#48814) huggingface/transformers
Generation no longer carries the attention mask forward when no padding is present, with a position_ids check covering the case where the mask is gone. The change also sets the length on the generation config to head off downstream issues.
[utils] Add MUSA support for Flash Attention 2 (#48612) huggingface/transformers
Flash Attention 2 availability detection now includes MUSA. The upsides are limited to MUSA users: the check simply reports FA2 as available on that backend where it previously did not.
fix: restore model customization reuse state huggingface/sagemaker-python-sdk
ModelBuilder reused Models and endpoints independently, so when build(reuse_resources=True) found an existing package-backed Model but deployment could not reuse an endpoint, the SDK took the endpoint-creation path without rebuilding the Model. State normally populated while building a customization Model, such as recipe-derived compute requirements and LoRA settings, is now restored in that path.
[OpenVINO] Add support for Qwen3-TTS model huggingface/optimum-intel
optimum-intel added OpenVINO export support for Qwen3-TTS, including int8 and int4 compression, with an export command for Qwen/Qwen3-TTS-12Hz-0.6B-Base and the OVModelForTextToSpeechSeq2Seq class for voice cloning.
Fixed vllm and vllm-omni tasks to correct routing logic for HF DLCs huggingface/sagemaker-python-sdk
The vLLM and vLLM-omni task sets used in Hugging Face DLC routing were wrong, selecting incorrect DLCs for models like google/gemma-4-E2B-it and other image-text-to-text models. Expect more movement here: two related changes were opened in the last day.
Support LTX2.3 huggingface/optimum-intel
An LTX2.3 text-to-audio-video pipeline joins optimum-intel's OpenVINO exports, and its landing commit documents that the exported LTX-2 sigma now ties to audio_timestep rather than the per-token timestep, which previously produced one modulation vector per video token and died in prompt_adaln.
Action items
- → Upgrade kernels past 0.17.0 if you build from source in an environment where it is installed huggingface/kernels [immediate]
