GitHub

7 articles in the TruthFoundry index. Each links out to the original.

• llama.cpp Optimizes GPU Context Creation to Prevent Unnecessary VRAM Allocation · llama.cpp developers fix a bug where device info loops unnecessarily allocated 550 MB of VRAM on CUDA systems.
• llama.cpp releases DeepseekV4 model fix for multi-sequence rollback · llama.cpp releases a pre-release update fixing DeepseekV4 model loading issues.
• llama.cpp Adds PAD_REFLECT_1D Operation to Vulkan Backend · The ggml-org team implemented the PAD_REFLECT_1D operation for the Vulkan backend in llama.cpp.
• llama.cpp Adds Dynamic Draft Context Sizing for Unified KV Cache · llama.cpp updates draft context sizing to match target context for unified KV cache support.
• llama.cpp Reverts Ubuntu ROCm CI Disable and Optimizes Build Caching · The llama.cpp project reverts a previous commit disabling ROCm CI jobs for Ubuntu and updates build caching strategies.
• llama.cpp Optimizes ggml Concat Operation with Row-Level Memory Copies · llama.cpp developers optimize the ggml library by replacing per-element memory copies with row-level copies in the concat operation.
• llama.cpp Releases Stable v0.2.0 with New Semantic Versioning · llama.cpp launches stable v0.2.0 release introducing consistent semantic versioning for downstream projects.