b10431

ggml-org/llama.cppb10431Aug 14, 2026by github-actions[bot]same day as b10430

What changed since b10430

medium risk

b10431 adds recurrent state rollback for ggml_ssm_scan (Nemotron) plus K>1 support, on top of b10430's virtual iGPU device support.

New features

  • Recurrent state rollback is now supported for ggml_ssm_scan, initially targeting Nemotron-style SSM models on CPU and CUDA.
  • ssm_scan now handles K > 1 across all backends rather than only the single-state case.
  • Graph selection is now driven at runtime by a fused-op capability query instead of a static backend probe.
  • Backends lacking SSM rollback support automatically fall back to CPU, and the memory->n_rs_seq > 1 case is handled when the op is unsupported.

Improvements

  • Replaced static CUDA detection with a runtime fused_op API check, making backend capability handling more portable.
  • Removed unnecessary memory manipulation now that rollback is natively supported on CPU.
  • Extended test-backend-ops coverage to exercise SSM rollback on both CPU and CUDA, with unsupported backends falling back.
  • Carried over b10430's virtual iGPU device support and comment documentation improvements.

Bug fixes

  • Fixed CI issues introduced during the SSM rollback work.
  • Removed an assert in favor of a runtime capability check, avoiding hard failures on backends without rollback support.

Upgrade notes

This is a routine build bump with no config or API migration required; existing backends and model files continue to work. SSM/Nemotron workloads may see different graph paths depending on whether the runtime reports fused SSM op support, so validate performance on non-CPU/CUDA backends where CPU fallback now applies. If you rely on custom ggml backends, confirm they correctly report fused-op capability to avoid unintended fallback.

Analyzed Oct 24, 2026

Upgrading?

Release notes summary

Implements recurrent state rollback support for SSM (State Space Model) operations in nemotron models, enhancing performance for certain model architectures.

Key Highlights

  • Recurrent state rollback for ggml_ssm_scan
  • Support for CPU and CUDA backends
  • Runtime API support for fused operations
  • Fallback mechanism when SSM rollback is not supported

Release Notes (excerpt)

<details open> ggml : recurrent state rollback for ggml_ssm_scan (#26623) * Initial changes for Recurrent state rollback for nemotron for cpu and cuda * Removing CPU RS rollback. Will enable it in subsequent PRs * addition of test case * Removing assert and calling runtime API to check if op is supported * removing extra API and updating the call sites for K * replace static cuda detection to runtime fused_op api * address review comments and fallback when SSM rollback not supprted…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories