cuda: fold the recurrent-state gather into the GATED_DELTA_NET kernel, per-context registry (+3.8% TG on Ada) - #220

cuda: fold the recurrent-state gather into the GATED_DELTA_NET kernel, per-context registry (+3.8% TG on Ada)#220

professorpalmer wants to merge 2 commits into

PrismML-Eng:prismPrismML-Eng/llama.cpp:prismfrom

professorpalmer:cuda-gdn-gather-fusionprofessorpalmer/llama.cpp-ada-ternary:cuda-gdn-gather-fusionCopy head branch name to clipboard