add other tma

Author: da.huo

src/turbomind/kernels/gemm/CMakeLists.txt

@@ -50,6 +50,7 @@ add_library(gemm2
         cast.cu
         unpack.cu
         context.cu
+        tma.cu
         tuner/cache_utils.cu
         tuner/measurer.cu
         tuner/sampler.cu
@@ -86,7 +87,7 @@ set_property(TARGET gemm2 PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS ON)
 
 if(GEMM2_ARCH_90_ENABLED)
     # SM90 kernels only compile for 90/90a; avoid building them for sm_100.
-    add_library(gemm2_sm90 STATIC ${GEMM2_KERNELS_SM90})
+    add_library(gemm2_sm90 STATIC ${GEMM2_KERNELS_SM90} )
     set_target_properties(gemm2_sm90 PROPERTIES
         CUDA_ARCHITECTURES "${_sm90_archs}"
         POSITION_INDEPENDENT_CODE ON