AI and ML model and collective updates

Activity in AI and ML focused on integrating a new model into a major library and refining distributed collective operations. Two technical proposals addressed architecture support and backend flexibility for training workloads.

DeepSeek-V4.1-Flash addition to Transformers

An issue in the Hugging Face Transformers repository proposes adding the DeepSeek-V4.1-Flash model under the deepseek_v41 identifier. The request includes support for custom attention and quantization. Readers tracking open model availability gain a concrete path for broader inference and fine-tuning access once implemented.

Uneven NCCL collectives in PyTorch c10d

A comment on a PyTorch c10d issue addresses uneven all_gather_into_tensor and reduce_scatter_tensor operations that can leverage NCCL symmetric memory. It suggests providing a backend-agnostic equivalent for these collectives. This matters for developers running distributed training who need consistent behavior across communication backends when tensor sizes are unbalanced.