b9079
Critical fix for reasoning models: llama.cpp rolls back buggy logit bias.
Deep Dive
The latest llama.cpp release (b9079) reverts the reasoning budget +inf logit bias (PR #22740). New builds are provided for macOS, iOS, Linux, Windows, Android, and openEuler.
Key Points
- Reverts reasoning budget +inf logit bias introduced in PR #22740
- Corrects logit values that were incorrectly set to infinity, skewing model outputs
- Available across all major platforms including macOS, Windows, Linux, Android, and iOS
Why It Matters
Essential fix for developers running reasoning-optimized LLMs locally, restoring accurate probability distributions and reliable outputs.