CYBER WATCH / EQUINOXE INTELLIGENCE
← All articles
b12257: CUDA: move graph and concurrent stream state out of common.cuh (#32001)
llama.cpp Releases · 2026-11-11 00:01:19
The structures ggml_cuda_graph, ggml_cuda_concurrent_event, and ggml_cuda_stream_context, previously defined in common.cuh, are now moved to graph.cuh. This prevents recompiling every CUDA translation unit when they change. The context constructor, cuda_graph(), and any_cuda_graph_*() functions are now defined in ggml-cuda.cu, and the stream context is managed via a unique_ptr. The function ggml_cuda_kernel_can_use_pdl is also defined once in ggml-cuda.cu, consolidating the PDL cache and incorporating the GGML_CUDA_PDL environment check.Read original source ↗