trunk/0c8b4a78ffbbce776adc0823e790158b26435f40: flex_gemm: grouped-main SwiGLU on the grouped_mm path (#196321)
- PyTorch 近 90 天出现 1128 次
- PyTorch 近 90 天第 1114 次发版
- 上一次:同一天稍早 · trunk/20fb7bd04e0b24cd4299fa34d767388daa6a26a4
发生了什么
Human Note Agent note The dense SwiGLU spellings (interleaved acc.view(m, -1, 2) + select , and chunked split over a column-major [2N, K].t() weight) now lower on the F.grouped_mm route, so torchtitan's MoE expert FFN silu(x @ w1) * (x @ w3) over ragged experts is one grouped FlexGEMM with the same epilogue text a dense user writes. QuACK's GroupedMainStore already accepted varlen-M; only the lowering guard rejected…
摘要按规则整理自下方来源原文