diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/CMakeLists.txt b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/CMakeLists.txt
index daa5a3e1d974..5731ed4f0549 100644
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/CMakeLists.txt
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/CMakeLists.txt
@@ -49,7 +49,15 @@ configure_file(
 file(GLOB_RECURSE SRC_CPP *.cpp)
 file(GLOB_RECURSE SRC_CU *.cu)
 
-filter_source_cuda_architectures(SOURCE_LIST SRC_CPP ARCHS 100 103 107 100f)
+# Archs listed here are pruned when the build does not enable them. Sm100f has
+# no context FMHA kernel, so Rubin needs the prebuilt Sm107a cubins kept
+# unconditionally: 107 is omitted below, and the globally-set EXCLUDE_SM_107
+# that would drop their metadata rows is cleared for this directory.
+set(FMHA_PRUNABLE_ARCHS 100 103 100f)
+remove_definitions(-DEXCLUDE_SM_107)
+
+filter_source_cuda_architectures(SOURCE_LIST SRC_CPP
+                                 ARCHS ${FMHA_PRUNABLE_ARCHS})
 
 add_library(trtllm_gen_fmha_interface INTERFACE)
 target_include_directories(
@@ -65,7 +73,8 @@ target_include_directories(
 
 include(tllm_cubin_archive)
 tllm_add_cubin_archive_sources(
-  trtllm_gen_fmha "${CMAKE_CURRENT_SOURCE_DIR}/cubin" ARCHS 100 103 107 100f)
+  trtllm_gen_fmha "${CMAKE_CURRENT_SOURCE_DIR}/cubin"
+  ARCHS ${FMHA_PRUNABLE_ARCHS})
 set(TRTLLM_FMHA_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}")
 target_compile_definitions(
   trtllm_gen_fmha