AI and ML library hardware proposals

Proposals for hardware-oriented changes appeared in two widely used open source machine learning codebases. Work centered on quantized cache support in llama.cpp and backend availability checks in transformers.

FP4 E2M1 KV cache work in llama.cpp

A single-message pull request update in the ggml-org/llama.cpp repository proposes FP4 E2M1 KV cache support. The change targets RTX 5090 and Blackwell GPUs. Developers tracking inference efficiency on current NVIDIA hardware have a concrete optimization path under discussion.

MUSA backend request for flash linear attention

A feature request in the huggingface/transformers repository asks to extend the flash-linear-attention availability check to the MUSA backend. The single-message proposal covers Moore Threads GPU hardware support. Users of alternative accelerator vendors gain a potential path to the same attention kernels already gated for other backends.