infrastructure
Llama.cpp caps working memory size to prevent large tensor RAM loads.
The update caps working memory size to prevent large tensors from being loaded into RAM.
1 sources · 1 verified claimsAI knowledge entity
Llama Cpp
One evolving timeline of verified developments, without duplicating the underlying truth across pages.
10 verified events1 verified sourceUpdated 27 Aug 2026
Verified timeline
Developments connected to this entity
infrastructure
Two new bin kernels added to OpenCL in llama.cpp
Two new bin kernels added for optimized performance on various platforms.
1 sources · 1 verified claimsinfrastructure
Llama adds TENSOR_READ_LAZY to model loader
Llama adds TENSOR_READ_LAZY and TENSOR_GET_ROW_LAZY for improved model loading.
1 sources · 1 verified claimsinfrastructure
Llama-server and llama-cli gain benchmark-only synthetic speculative acceptance.
Llama-server and llama-cli now include benchmark-only synthetic speculative acceptance options.
1 sources · 1 verified claimsinfrastructure
Apple patches Metal memory leaks in macOS and iOS.
A memory leak fix for Metal has been implemented in the latest update. The fix addresses missing autoreleasepools.
1 sources · 1 verified claimsinfrastructure
Llama adds CPU-offload option for dense FFN weights
The --n-cpu-ffn option is added to offload dense FFN weights of the first N layers on CPU.
1 sources · 1 verified claimsinfrastructure
MiniMax-01 graph simplified in model update.
The MiniMax-01 graph model was simplified by Stanisław Szymczyk. It supports macOS Apple Silicon, Ubuntu, and Windows.
1 sources · 1 verified claimsinfrastructure
Llama.cpp adds video CLI arguments in latest update
The new CLI arguments include video options for macOS, Linux, Windows, and Android.
1 sources · 1 verified claimsinfrastructure
KleidiAI macOS Apple Silicon support disabled in latest llama.cpp update
Vulkan's warptiles currently assume specific warp sizes. The issue is tracked on GitHub.
1 sources · 1 verified claimsinfrastructure
Llama adds token ID tracking to key-value cell.
Token ID tracking added to KV cell in llama. Supports macOS, Linux, Windows, and Android.
1 sources · 1 verified claims