[gemma4_31b][cuda] length-aware bf16 global attention + head_dim-agno…#20506

Gasoonjia wants to merge 1 commit into

gemma4_31b-cuda-decode-speeduppytorch/executorch:gemma4_31b-cuda-decode-speedupfrom

gemma4_31b-cuda-attn-perf-gitpytorch/executorch:gemma4_31b-cuda-attn-perf-gitCopy head branch name to clipboard