perf(cuda): adaptive split-KV sizing for attention_row decode (+4-5% over fixed-256; +70% deep-ctx) - #1350

perf(cuda): adaptive split-KV sizing for attention_row decode (+4-5% over fixed-256; +70% deep-ctx)#1350

justinchuby merged 1 commit into

mainjustinchuby/onnx-genai:mainfrom

squad/attention-splitkv-adaptivejustinchuby/onnx-genai:squad/attention-splitkv-adaptiveCopy head branch name to clipboard