trunk/610fee8d845047496bd0713e7adff4d79d08f137: [ROCm] Fix CK SDPA flash attention on gfx1200/gfx1201 (RDNA4) (#188114)
- PyTorch 近 90 天出现 180 次
- PyTorch 近 90 天第 178 次发版
- 上一次：同一天稍早 · trunk/29343e76e2342c777bb257a270bae1d805996700: Support non-builtin constexpr objects in user Triton kernels (#189692)
发生了什么
The CK Tile FMHA generator already compiles working kernels for gfx12 (--targets=gfx1201 passes through to the kernel template), but two guards prevented them from being used on RDNA4 hardware: launch_kernel_pt.hpp: kentry_pt kernel entry stubs added in PR #144777 ([ROCm] CK SDPA - Move arch check to CK patch) had a hard-coded whitelist of gfx90a/gfx942/gfx950 (updated for gfx950 in PR #155103 ). gfx1200/gfx1201 wer…
摘要按规则整理自下方来源原文