llama.cpp b10431 adds Nemotron SSM rollback for faster CPU/CUDA inference

ggml-org's llama.cpp, the widely used open-source C++ LLM inference engine with 124k stars and 21.7k forks, released version b10431 on August 14. The release introduces recurrent state rollback for ggml_ssm_scan (PR #26623), a mechanism that improves state space model (SSM) inference by efficiently handling recurrent state rollback during sequential scans. This initial implementation specifically targets Nemotron, NVIDIA's SSM-based model architecture, with CPU and CUDA backends supported out of the box. The contribution was led by developers including Georgi Gerganov, the project's creator, and Gaurav Garg from NVIDIA, with fixes for CI issues and comprehensive test-backend-ops coverage for both CPU and CUDA.

Beyond the Nemotron rollout, b10431 extends support for K > 1 in ssm_scan across all backends, a significant improvement for multi-sequence state space model processing. The implementation uses a runtime API check to determine if the fused SSM operation is supported on a given backend, falling back to CPU when needed. This eliminates static probing and enables a single unified graph selection path. For unsupported backends, the system gracefully defaults to CPU, ensuring reliability across diverse hardware including Vulkan, ROCm, SYCL, OpenCL, and OpenVINO. The release also removes memory manipulation overhead now that CPU supports rollback natively, and includes Windows, macOS, Linux, Android, and openEuler binaries. This update makes llama.cpp more robust for state space models, benefiting developers running Nemotron and similar architectures locally with optimized performance.