PyTorch rewrite breaks vLLM, llama.cpp RAM codec

A PyTorch functionalization rewrite is breaking a vLLM fusion path, while llama.cpp and ROCm Windows work continue on memory and build enablement. Smaller cache and precompile proposals landed in Transformers and PyTorch with little discussion.

PyTorch split rewrite regresses vLLM QK-norm and RoPE fusion

PyTorch PR194037 replaced split_with_sizes with slices under functionalization, and issue 197358 reports that the change breaks vLLM's QK-norm and RoPE fusion pass. The regression sits at the intersection of core tensor rewrites and production LLM serving stacks. Operators and vLLM maintainers need a fix or workaround before the functionalization path can ship cleanly for those kernels.

Opt-in codec to cut llama.cpp recurrent checkpoint RAM

An opening issue in ggml-org/llama.cpp proposes an opt-in codec for recurrent-state context checkpoints that would roughly halve host RAM use in the server. The feature would stay off by default. Server operators running long recurrent contexts care because host memory, not just GPU VRAM, often limits concurrent sessions.

ROCm Windows path to enable CK SDPA

A PyTorch PR sketches CMake, linker, and macro fixes plus non-MI3xx fav_v3 decoupling and RDNA4 gfx120x enablement so CK SDPA can build on Windows ROCm. It reverses an earlier guard that blocked the path after build failures. Windows ROCm users gain a clearer route to the same SDPA backend already used on Linux.

Static cache size relaxation for encoder-decoder export

A Hugging Face Transformers feature request asks to allow static caches larger than sequence length or batch size for encoder-decoder models. The change targets ExecuTorch export constraints that currently demand exact sizes. Export pipelines that pad or reuse oversized caches would no longer need bespoke workarounds.

BarkModel history prompt exposure and other small PRs

A lone Transformers request seeks to return output_full or history_prompt from BarkModel.generate for callers that need the full generation state. Separate low-activity PyTorch PRs cover keeping a legacy ProcessGroupNCCL dump-signal responder alive during communicator teardown, adding capture and load APIs for precompile on-disk artifacts, and an FP8 backend for QuantizedCache that is still blocked on a CI security gate.