gfx928

cmake/external/flashattn.cmake

@@ -33,7 +33,7 @@ if(WITH_ROCM)
     set(FLASHATTN_C_FLAGS ${CMAKE_C_FLAGS})
     set(FLASHATTN_C_FLAGS_DEBUG ${CMAKE_C_FLAGS_DEBUG})
     set(FLASHATTN_C_FLAGS_RELEASE ${CMAKE_C_FLAGS_RELEASE})
-    set(FLASHATTN_CXX_FLAGS "${CMAKE_CXX_FLAGS} -w -fPIC -O3 -std=c++17 -D__HIP_PLATFORM_HCC__=1 --offload-arch=gfx906 --offload-arch=gfx926 -D__gfx90a__")
+    set(FLASHATTN_CXX_FLAGS "${CMAKE_CXX_FLAGS} -w -fPIC -O3 -std=c++17 -D__HIP_PLATFORM_HCC__=1 --offload-arch=gfx906 --offload-arch=gfx926 --offload-arch=gfx928 -D__gfx90a__")
     set(FLASHATTN_CXX_FLAGS_RELEASE ${CMAKE_CXX_FLAGS_RELEASE})
     set(FLASHATTN_CXX_FLAGS_DEBUG ${CMAKE_CXX_FLAGS_DEBUG})
 

cmake/hip.cmake

@@ -144,9 +144,11 @@ set(HIP_CLANG_FLAGS ${HIP_CXX_FLAGS})
 list(APPEND HIP_HCC_FLAGS -fno-gpu-rdc)
 list(APPEND HIP_HCC_FLAGS --offload-arch=gfx906)
 list(APPEND HIP_HCC_FLAGS --offload-arch=gfx926)
+list(APPEND HIP_HCC_FLAGS --offload-arch=gfx928)
 list(APPEND HIP_CLANG_FLAGS -fno-gpu-rdc)
 list(APPEND HIP_CLANG_FLAGS --offload-arch=gfx906)
 list(APPEND HIP_CLANG_FLAGS --offload-arch=gfx926)
+list(APPEND HIP_CLANG_FLAGS --offload-arch=gfx928)
 
 if(HIP_COMPILER STREQUAL clang)
   set(hip_library_name amdhip64)