b10258

ggml-org/llama.cppb10258Aug 4, 2026by github-actions[bot]same day as b10256

What changed since b10256

low risk

Internal sampler refactor moves n_vocab into penalty_sampler; no user-facing behavior changes between b10256 and b10258.

Improvements

  • Sampler internals reorganized so vocabulary size lives in the penalty sampler, matching how logit_bias and mirostat samplers already store it.
  • Cleaner, more consistent llama_sampler_data layout that should simplify future sampler work.

Upgrade notes

This is a purely internal refactor with no CLI, config, or API changes expected for typical users. Rebuild as usual; only downstream code that reaches into llama_sampler_data internals directly would need to adjust. The SYCL concat optimization from b10256 remains in place.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Refactors sampler data structure by moving vocabulary size (n_vocab) into the penalty_sampler struct.

Key Highlights

  • Refactor of llama_sampler_data
  • Moves n_vocab to penalty_sampler
  • Consistency with logit_bias and mirostat samplers

New Features

  • Sampler data structure refactor

Release Notes (excerpt)

<details open> llama : move n_vocab from llama_sampler_data to penalty_sampler (#26520) This matches how it is done for logit_bias and mirostat samplers, see https://github.com/ggml-org/llama.cpp/pull/25262#discussion_r3703951151 </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