viable/strict/1792232715: flex_gemm: support chunked GLU epilogue outputs (#201199)

发生了什么

Summary (human written only) Allow for the chunked form of interleaved weights Agent details (AI-generated) Supports concatenated gate/up gradients and saving the unchanged accumulator from a chunked GLU forward epilogue. Existing tile stores write into aligned halves of one allocation; no new CUDA kernel or vendored QuACK changes are needed. Scalar/dtype suffixes and the existing interleaved path are preserved. Tes…

摘要按规则整理自下方来源原文

来源