b10210

ggml-org/llama.cppb10210Jul 31, 2026by github-actions[bot]same day as b10209

What changed since b10209

low risk

Server-side fix corrects accepted-token accounting when speculative draft tokens must be replayed, replacing the prior CUDA Q2_0 kernel tweak as the only change.

Improvements

  • Speculative decoding token accounting in the server was reworked and renamed for clarity as part of the fix.
  • The previous release's CUDA Q2_0 __byte_perm element-extraction optimization remains in place.

Bug fixes

  • The server API now reports the correct number of accepted tokens when draft token replay is required during generation.
  • Speculative decoding paths that trigger replay no longer produce skewed acceptance counts in server responses.

Upgrade notes

This is a single-commit increment over b10209, so no config or model changes are needed. Upgrade if you rely on server-reported accepted-token statistics with speculative decoding enabled; otherwise the delta is negligible.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

This release fixes a bug in the server API where accepted tokens were not correctly counted when draft token replay was needed during generation.

Key Highlights

  • Corrects accepted tokens when draft token replay is needed
  • Fixes server API correctness

New Features

  • Server token replay fix

Release Notes (excerpt)

<details open> server: correct accepted tokens when need draft token replay (#26320) * spec: correct accepted tokens when need draft token replay * cont : naming --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - Ubuntu x64 (CPU) - Ubuntu arm64 (CPU) - Ubuntu s390x (CPU)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories