Community patch enables DeepSeek V4 Flash with 1M context on RTX 5090 via llama.cpp

A Reddit user published a patch for llama.cpp that allows running DeepSeek V4 Flash with full 1M token context locally on an RTX 5090, fixing a VRAM issue caused by missing DSA lightning indexer support. The patch is available as an upstream PR.