viable/strict/1791179152: Fix CUDA SDPA grid overflow for large batches and heads (#199594)

发生了什么

Human Note For batch and head dims that exceed the grid cap, we just launch multiple kernels used some features found in #177985 Agent Report Agent details Prepared with AI assistance. CUDA grid.y/grid.z cannot exceed 65,535 blocks. Chunk the memory-efficient forward/backward launch grids while retaining full tensor dimensions and logical batch/head offsets. Kernels write directly into final outputs, preserving GQA…

摘要按规则整理自下方来源原文

来源