Local Ai
b9292
Release b9292 of llama.cpp fixes a memory leak in the server context where speculative decoder, draft context, and draft model were not properly freed during destroy(), causing VRAM leaks on sleep/res
Release b9292 of llama.cpp fixes a memory leak in the server context where speculative decoder, draft context, and draft model were not properly freed during destroy(), causing VRAM leaks on sleep/resume cycles. The fix explicitly resets these resources before the main model to ensure proper cleanup order and prevent out-of-memory errors. Additional updates include refactoring Carbon-3B tokenizer changes to align with llama.cpp conventions.
Source: llama.cpp Releases | 2026-05-22