llama.cpp b10005 fixes critical DeepseekV4 sequence handling bugs
New patch resolves sequence removal and copying issues for DeepseekV4 models
The latest patch release of llama.cpp, b10005, addresses critical bugs in DeepseekV4 model support. Specifically, it fixes the seq_rm function, which handles sequence removal, and implements a proper seq_cp for creating actual updated context. These fixes are essential for maintaining coherent context windows during multi-turn interactions and batch processing, which are common in local LLM deployments.
The b10005 release is available for all major platforms, including macOS (Apple Silicon and Intel), Linux (x64, ARM, s390x), Windows (x64, ARM), iOS, and Android. The GitHub release includes builds with various acceleration backends like CUDA, Vulkan, ROCm, OpenVINO, and SYCL. Users running DeepseekV4 models on llama.cpp should update to this version to avoid errors in sequence management.
- Fixes seq_rm (sequence removal) bug for DeepseekV4 models, improving context management
- Implements proper seq_cp (sequence copy) to create actual updated context for ongoing conversations
- Available across all major platforms including macOS, Linux, Windows, iOS, and Android with multiple acceleration backends
Why It Matters
Ensures reliable multi-turn conversations with DeepseekV4 models in local LLM deployments, critical for developers running open-source AI.