[ROCm][Perf] W4A16: speed up skinny GEMM decode, including speculative batches (#57199)

Signed-off-by: Robert Esclapez Garcia <robert.garcia@amd.com>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Co-authored-by: Douglas Lehr <91553416+dllehr-amd@users.noreply.github.com>
This commit is contained in:
Robert Esclapez
2026-10-13 17:21:30 +00:00
committed by GitHub
co-authored by Claude Opus 4.7 Douglas Lehr
parent 7fad7db330
commit e83222396f
4 changed files with 603 additions and 117 deletions
+5
View File
@@ -218,6 +218,7 @@ _ON_GFX1X = any(arch in _GCN_ARCH for arch in ["gfx11", "gfx12"])
_ON_GFX11 = "gfx11" in _GCN_ARCH
_ON_GFX1100 = "gfx1100" in _GCN_ARCH
_ON_GFX1151 = "gfx1151" in _GCN_ARCH
_ON_GFX115X = "gfx115" in _GCN_ARCH
_ON_GFX12X = any(arch in _GCN_ARCH for arch in ["gfx12"])
_ON_MI3XX = any(arch in _GCN_ARCH for arch in ["gfx942", "gfx950"])
_ON_GFX9 = any(arch in _GCN_ARCH for arch in ["gfx90a", "gfx942", "gfx950"])
@@ -319,6 +320,10 @@ def on_gfx1151() -> bool:
return _ON_GFX1151
def on_gfx115x() -> bool:
return _ON_GFX115X
def on_gfx12x() -> bool:
return _ON_GFX12X and not _ON_CDNA