Hugging Face AI Updates: July 28, 2026

1. huggingface_hub v1.25.0 Adds Job Auto-Naming and Cache Consistency Warnings

Hugging Face published version 1.25.0 of huggingface_hub, its Python client for the Hub. Jobs now receive an automatic name derived from the Docker image plus a command hash when no explicit name is provided, with server-side character rules that replace colon, slash, and period in image tags with hyphens so reruns and grouping stay consistent. The release adds a timeout parameter to safetensors metadata methods, surfaces warnings when it detects cache inconsistencies, and fixes progress bars during downloads along with bug corrections affecting Windows compatibility, repository card handling, and file system operations. Source

2. TRL v1.9.1 Fixes Liger Kernel Crash and Loss Normalization

Hugging Face shipped TRL v1.9.1, a patch release for its reinforcement learning and post-training library. The update fixes a crash when training with the Liger kernel on pre-Ampere GPUs and temporarily pins liger-kernel below 0.8.1 to keep continuous integration stable. It also corrects DAPO, CISPO, and VESPO loss normalization when steps_per_generation differs from gradient_accumulation_steps, ensuring accurate loss calculations across training configurations. Additional fixes cover vLLM server initialization, AsyncGRPOTrainer metrics, DeepSpeed CPU offload crashes, and Gemma4 response template support. Source