Hugging Face Transformers v5.22.2 patches NCCL device handling and Jinja sandboxing

Hugging Face released Transformers v5.22.2, a security and distributed-training patch for the widely-used Apache-2.0 machine-learning library. The release contains two targeted fixes: one hardens Jinja2 template sandboxing to reduce code-injection risks in model card and prompt templating, and another corrects distributed initialization logic so that device_id is only configured for the NCCL backend, preventing potential runtime errors with other backends like Gloo or MPI.

This is a maintenance release with 144 commits landing on the main branch since publication, indicating active ongoing development toward the next minor version. Users running distributed training across multiple GPUs or relying on automated template rendering in production pipelines should prioritize this update.

Security and distributed training fixes in v5.22.2

• Jinja sandboxing hardening: The patch strengthens isolation for Jinja2 template execution, which Transformers uses for model card rendering, chat templates, and other dynamic text generation; this mitigates sandbox escape vectors that could allow arbitrary code execution through malicious template input.
• NCCL-specific device_id handling: The distributed training fix ensures device_id is only passed when the NCCL backend is active, resolving incorrect behavior where the parameter was previously set unconditionally and could conflict with non-NCCL backends such as Gloo or MPI during multi-node or multi-GPU initialization.
• Release scope: Only two merged pull requests are highlighted for this patch, suggesting a focused emergency-fix release rather than a broader feature update, with 144 subsequent commits already queued on main indicating accelerated development cadence.

Why it matters

For the thousands of downstream projects and commercial services built on Transformers, the Jinja sandboxing fix closes a template-injection surface that is particularly exposed in automated ML pipelines and hosted inference APIs. The NCCL device_id correction prevents subtle distributed-training failures on heterogeneous clusters where operators mix backend types, reducing debugging overhead for researchers and MLOps engineers running large-scale fine-tuning jobs.