llama.cpp AMD speedups and model adds, PyTorch API notes

llama.cpp saw a Vulkan matmul path for AMD GPUs and a new MoE model PR, while PyTorch threads floated autograd and Inductor design ideas. The day favored inference runtime work over framework core changes.

Vulkan int8 coopmat1 matmul for AMD RDNA in llama.cpp

A pull request in ggml-org/llama.cpp adds an int8 coopmat1 matmul implementation over Vulkan for AMD RDNA3 and RDNA4. A user reports large inference speedups on an AMD 7900XTX from the change. Developers running local models on recent AMD cards gain a concrete path to higher throughput.

GLM-5.3-Flash MoE support in llama.cpp

A PR proposes adding the GLM-5-Next (GLM-5.3-Flash) MoE model to llama.cpp. The short discussion covers performance tuning and image support issues. Model runners tracking open backends will see another contemporary architecture entering the runtime.

Decorator-style API for PyTorch custom autograd

A single-message feature request in pytorch/pytorch proposes a functional decorator API (torch.autograd.op) to simplify writing custom torch.autograd.Function operators. The idea targets reduced boilerplate for user-defined differentiable ops. Contributors who extend autodiff would have a lighter registration surface if adopted.

Inductor LDS retention prototype via TritonScheduling

A two-message thread on a TLX generic cross-phase LDS retention prototype in pytorch/pytorch argues for subclassing TritonScheduling instead of adding generic Inductor hooks. The exchange stays at the level of how to structure the prototype. Compiler and GPU kernel authors watching Inductor internals can note the preferred extension style.