npx @buildinternet/releases get hugging-faceNew chunked loss option for SFT reduces peak activation memory by processing tokens in checkpointed chunks, enabling longer sequences without OOM errors. Includes new DPO loss variant, OpenReward adapter, Cohere Cohere2 Gemma 3 Qwen3 Qwen2.5 chat templates; fixes CLI wildcard uploads and click 8.5 compatibility.
New pipelines include MiniMax-H3 (video+audio), MiniMax Music 3, Stable Audio 3, LTX-2.5, Wan-Animate-2, and JoyAI-Image-Edit-Plus. Modular Diffusers is no longer experimental, and minimal tensor-parallel support is added.
Fixed DFlash candidate token device mismatch with device_map='auto', aligned logit distributions for sampling-based candidate generators, fixed MTP config when mlp_layer_types is absent, and added a fallback from Lanczos to bicubic filtering on CUDA so images process on accelerator. Also fixed gemma4 video device placement.
The new hf endpoints hardware command lists valid hardware combinations with pricing and quota, mirrored in the SDK via list_inference_endpoints_hardware(). custom_image now accepts engine-specific container types (keyed by engine name) with new --engine, --tensor-parallel-size, and --data-parallel-size flags for deploy and update, while INFERENCE_ENDPOINT_IMAGE_KEYS is removed.
DistillationTrainer and DistillationConfig move from trl.experimental.distillation to the top-level trl package, with the old path emitting a FutureWarning until removal in v2.0.0. AsyncGRPOTrainer gains an experimental loop-owning path for training external agents like opencode, plus VLM support in DistillationTrainer and new DiffusionGemma SFT example.
Adds support for Meta's Muse Glimmer 30B multimodal model plus GraniteSWA, GraniteMoeSWA, A.X-K1/K2, and Cosmos3 Edge. Kernels for linear attention models (Mamba, GDN, Conv-only) are now opt-in rather than mandatory, cache cropping only accepts negative offsets, and T5 family gains SDPA support with a possible default attention change.
The hf-cli skill for AI agents is now installed globally by default by the standalone installers (bash and PowerShell), refreshed on hf update, with a once-daily hint if missing — silenced by HF_HUB_DISABLE_UPDATE_CHECK=1. --container-command and --container-args no longer require --custom-image for Inference Endpoints and can be changed after deploy via hf endpoints update; --health-route and --port still require a custom image. Baseten joins as an inference provider for the conversational task, plus a bucket prefix collision fix in HfFileSystem and a fix preventing double deletion of snapshot files.
The lerobot.types module was renamed to lerobot.lerobot_types. Additions include gradient checkpointing for diffusion models, slice support in LeRobotDataset, manual exposure/gain/white balance for RealSense cameras, private HF Hub dataset support, and a --no_stamp option for recording.
HfApi.resolve_revision and ResolvedRevision let libraries pin a revision once and guarantee all files come from the same commit. Two security fixes: downloads now reject absolute, UNC, and traversal filenames on all platforms, and the sandbox bootstrap no longer injects the HF token. Resource groups are now supported for Jobs and Collections, and the CLI surfaces job names as a first-class field.
Fixed a symlink-following arbitrary file write in archive extraction and a path traversal via metadata file_name in folder-based builders. Also fixed batch(by_column=...) crashing after shard/shuffle/split, CSV loader issues on pandas 2.0-2.2, and more than 20 other bugs across Arrow, JSON, and dataset card handling.
Jobs now get an automatic name derived from the Docker image plus a command hash, making reruns and grouping easier. Also adds timeout parameter to safetensors metadata methods, updates file-count progress bars on download completion, and warns on cache inconsistency.
Fixes a crash when training with Liger kernel on pre-Ampere (Fermi, Kepler, Maxwell, Pascal) GPUs and corrects DAPO, CISPO, and VESPO loss normalization when steps_per_generation differs from gradient_accumulation_steps. Also fixes vLLM server-mode communicator initialization, queue wait time metric in AsyncGRPOTrainer, and prepare_deepspeed crash with CPU offload optimizer.