viable/strict/1788331551: [ROCm][CK] Remove host-side FMHA sequence padding workaround (#195498)
- PyTorch 近 90 天出现 417 次
- PyTorch 近 90 天第 414 次发版
- 上一次：同一天稍早 · ciflow/trunk/195663: Update
发生了什么
Summary: Context #187152 added a gfx950 host-side pad/copy stopgap after unpadded CK FMHA TRLOAD instances faulted on tile-unaligned Q/K sequence lengths. ROCm/rocm-libraries#11519 guards those generated dispatch arms so unaligned requests select CK's existing tail-safe kPadSeqLenQ/K instances. What Remove the external Q/K/V sequence padding, O/LSE scratch allocation, and output copies from the PyTorch CK FMHA wrapp…
摘要按规则整理自下方来源原文