Hugging Face

TRL v1.5 and Transformers v5.9 shipped this week, extending a run of library updates across the ecosystem.

TRL's SFT trainer cut peak VRAM by up to 50% with a new loss_type="chunked_nll" option that avoids materializing the full [batch × seq × vocab] logits tensor. The same chunked path works for VLMs.¹ v1.5 followed with training chat templates for Phi-3.5, Qwen3-VL, and Qwen3.5, plus a fix for exponential-backtracking regex hangs in GRPO response parsing. Trainer telemetry (opt-out) now tracks which trainer classes and distributed backends are in use.²

Transformers added Cohere2Moe, HRM-Text, and Parakeet TDT in v5.9, with tensor parallelism support in continuous batching. A breaking change in SAM3 / EdgeTAM now expects full text embeddings rather than pooler outputs.³

huggingface_hub gained multimodal inference and tighter token security — Together AI's Inference Providers now accept five new task types including feature extraction, text-to-speech, and text-to-video. Token files are now written at 0o600 permissions (matching GitHub CLI and AWS CLI conventions).⁴ The hf:// URI parsing was consolidated onto a single parse_hf_uri helper; single-segment repo IDs (e.g. gpt2 without a namespace) are no longer accepted on HfFileSystem paths.⁵ Earlier releases in the window added region selection for buckets and repos, Space secrets management from the CLI, and Homebrew-aware version checks.⁶

smolagents v1.25 patched remote executor security — a high-impact vulnerability in remote executors was fixed, LocalPythonExecutor documentation now clarifies it is not a sandboxing tool, and the WasmExecutor now binds to loopback only.⁷

Platform additions include Kernels Hub (precompiled GPU kernels with 1.7–2.5× speedups), hf-mount for attaching Hub storage as a local filesystem, per-Space /agents.md endpoints for coding agents, and paid over-quota credits for ZeroGPU.⁸⁹¹⁰¹¹