llama.cpp NCCL work and Transformers attention PR

Activity in open ML tooling stayed narrow, with two low-volume pull requests touching parallelism and attention kernels. Neither thread advanced beyond initial review requests.

NCCL tensor-parallel support proposed for llama.cpp

A pull request in ggml-org/llama.cpp seeks to add NCCL support for tensor parallelism. Discussion remained minimal, limited to a request for benchmarks. The change would matter for multi-GPU inference scaling in the widely used local LLM runtime.

SDPA and Flash Attention for Decision Transformer

A new pull request in huggingface/transformers aims to enable SDPA and Flash Attention for the Decision Transformer model. A bot posted a request for maintainer CI approval. The work targets more efficient attention computation in that architecture.