$ the-wire · showcase
Diffusers 0.42.0 requires Python 3.11, lerobot cuts sidecar memory 91%
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
Diffusers shipped 0.42.0 with FLUX.3 Image Klein, Bria Fibo 2.0 and Kandinsky 6 but raised the floor to Python 3.11, while lerobot landed three separate policy and data-loading gains and dataset-viewer moved to huggingface_hub 2.
Diffusers 0.42.0: FLUX.3, Fibo 2.0, Kandinsky 6, and more huggingface/diffusers
The release also brings SANA-WM and RFD4-Proteína, new adapter support and modular pipelines, but the line that matters for existing environments is hard: "Python 3.11 or newer is now required." FLUX.3 Image Klein covers text-to-image and editing with one or more reference images, base and step-distilled, both through ModularPipeline.
perf(streaming): cut sidecar build memory by 91% huggingface/lerobot
The sidecar index that records frame times, durations and in-file frame offsets (introduced for selective episode fetching) now builds and caches at 91% less memory, which makes training on selected episodes without downloading every video practical on smaller hosts.
Use huggingface_hub 2 huggingface/dataset-viewer
The move to huggingface_hub 2 pulls in the files-listing optimization from huggingface_hub and a datasets bump; the stated purpose is datasets with very large file counts like objaverse, tying the viewer work to the current 3D data effort.
feat(fastwam): add Faster-WAM (2.5-5.3x faster policy inference on A100) huggingface/lerobot
An opt-in DoT variant of the fastwam policy implementing the three world-model-centric design choices from the Faster-WAM paper, enabled via FastWAMConfig with action_expert_layers=1, lite_kv_fusion, retracted_1d_rope and world_model_only_conditioning; the latency reduction is measured on A100.
feat(policies): add LingBot-VLA 2.0 huggingface/lerobot
LingBot-VLA 2.0 lands as policy.type=lingbot_vla_v2, a Qwen3-VL-4B backbone with a sparse-MoE Qwen2 action expert trained with flow matching on a unified 55-D state/action space, post-trained under lerobot-train and evaluated through lerobot-eval. Elsewhere: diffusers dropped six expired non_ema_revision blocks whose helper now raises, so plain --revision=main no longer crashes with a TypeError...
Action items
- → Move diffusers environments to Python 3.11 before upgrading to 0.42.0 huggingface/diffusers [plan]
