diff --git a/cmake/onnxruntime_cuda_source_filters.cmake b/cmake/onnxruntime_cuda_source_filters.cmake
index e65b58a2f3f23..2963a3e697469 100644
--- a/cmake/onnxruntime_cuda_source_filters.cmake
+++ b/cmake/onnxruntime_cuda_source_filters.cmake
@@ -3,7 +3,7 @@
 
 # Shared filtering logic for CUDA contrib ops .cu source lists.
 # Both the main CUDA provider and the plugin EP build use identical filtering
-# rules for flash attention (quick build) and MoE GEMM FP4/FP8 kernels.
+# rules for flash attention, XQA, and MoE GEMM FP4/FP8 kernels.
 #
 # Usage:
 #   onnxruntime_filter_cuda_cu_sources(<list_variable_name>)
@@ -143,6 +143,25 @@ function(onnxruntime_extract_flash_attention_sources CU_SRC_LIST)
   set("${_FA_FLASH_SOURCES}" "${_flash_srcs}" PARENT_SCOPE)
 endfunction()
 
+# Extract XQA CUDA source files into a separate list for SM80+ compilation.
+function(onnxruntime_extract_xqa_sources CU_SRC_LIST)
+  cmake_parse_arguments(PARSE_ARGV 1 _XQA "" "XQA_SOURCES" "")
+
+  set(_list "${${CU_SRC_LIST}}")
+  set(_xqa_srcs)
+  foreach(_src IN LISTS _list)
+    if(_src MATCHES "/bert/xqa/.*\\.cu$")
+      list(APPEND _xqa_srcs "${_src}")
+    endif()
+  endforeach()
+  if(_xqa_srcs)
+    list(REMOVE_ITEM _list ${_xqa_srcs})
+  endif()
+
+  set("${CU_SRC_LIST}" "${_list}" PARENT_SCOPE)
+  set("${_XQA_XQA_SOURCES}" "${_xqa_srcs}" PARENT_SCOPE)
+endfunction()
+
 # Extract LLM CUDA source files into separate lists for per-architecture compilation.
 # The LLM directory (contrib_ops/cuda/llm/) contains kernels with minimum SM75 support
 # (fpA_intB_gemv/gemm enforce arch >= 75). SM90-specific launchers (fpA_intB_gemm
diff --git a/cmake/onnxruntime_providers_cuda.cmake b/cmake/onnxruntime_providers_cuda.cmake
index 52eb24579cf8c..53a89b2ca5a0c 100644
--- a/cmake/onnxruntime_providers_cuda.cmake
+++ b/cmake/onnxruntime_providers_cuda.cmake
@@ -83,6 +83,9 @@
   onnxruntime_extract_flash_attention_sources(onnxruntime_cuda_contrib_ops_cu_srcs
     FLASH_SOURCES onnxruntime_cuda_flash_attention_srcs
   )
+  onnxruntime_extract_xqa_sources(onnxruntime_cuda_contrib_ops_cu_srcs
+    XQA_SOURCES onnxruntime_cuda_xqa_srcs
+  )
   onnxruntime_extract_llm_sources(onnxruntime_cuda_contrib_ops_cu_srcs
     LLM_SOURCES onnxruntime_cuda_llm_srcs
     LLM_SM90_SOURCES onnxruntime_cuda_llm_sm90_srcs
@@ -524,6 +527,26 @@
       endif()
     endif()
 
+    # XQA kernels require SM80+. Compiling them with mixed SM75/SM80+ architectures
+    # causes CUDA 13.3 to emit host references to kernels omitted from the SM75 pass.
+    if(onnxruntime_cuda_xqa_srcs)
+      onnxruntime_filter_cuda_archs(_ort_xqa_cuda_architectures MIN_SM 80)
+      if(_ort_xqa_cuda_architectures)
+        onnxruntime_add_cuda_object_library(
+          NAME onnxruntime_providers_cuda_xqa
+          PARENT onnxruntime_providers_cuda
+          CUDA_ARCHITECTURES "${_ort_xqa_cuda_architectures}"
+          NVCC_THREADS "${onnxruntime_NVCC_THREADS}"
+          SOURCES ${onnxruntime_cuda_xqa_srcs})
+      else()
+        if(TARGET onnxruntime_providers_cuda_obj)
+          target_sources(onnxruntime_providers_cuda_obj PRIVATE ${onnxruntime_cuda_xqa_srcs})
+        else()
+          target_sources(onnxruntime_providers_cuda PRIVATE ${onnxruntime_cuda_xqa_srcs})
+        endif()
+      endif()
+    endif()
+
     if(NOT onnxruntime_DISABLE_CONTRIB_OPS)
       # SM90 TMA warp-specialized files use SM90-specific collective operations.
       # Compile at exactly 90a-real: SM120+ GPUs run SM90 native code via forward compat.
diff --git a/cmake/onnxruntime_providers_cuda_plugin.cmake b/cmake/onnxruntime_providers_cuda_plugin.cmake
index 9e940b0103bb8..99f1ced14ebd1 100644
--- a/cmake/onnxruntime_providers_cuda_plugin.cmake
+++ b/cmake/onnxruntime_providers_cuda_plugin.cmake
@@ -120,6 +120,9 @@ onnxruntime_extract_sm_specific_cuda_sources(CUDA_PLUGIN_EP_CU_SRCS
 onnxruntime_extract_flash_attention_sources(CUDA_PLUGIN_EP_CU_SRCS
   FLASH_SOURCES _cuda_plugin_flash_attention_srcs
 )
+onnxruntime_extract_xqa_sources(CUDA_PLUGIN_EP_CU_SRCS
+  XQA_SOURCES _cuda_plugin_xqa_srcs
+)
 onnxruntime_extract_llm_sources(CUDA_PLUGIN_EP_CU_SRCS
   LLM_SOURCES _cuda_plugin_llm_srcs
   LLM_SM90_SOURCES _cuda_plugin_llm_sm90_srcs
@@ -355,6 +358,21 @@ if(_cuda_plugin_flash_attention_srcs)
   endif()
 endif()
 
+if(_cuda_plugin_xqa_srcs)
+  onnxruntime_filter_cuda_archs(_plugin_xqa_cuda_architectures MIN_SM 80)
+  if(_plugin_xqa_cuda_architectures)
+    onnxruntime_add_cuda_plugin_object_library(
+      NAME onnxruntime_providers_cuda_plugin_xqa
+      PARENT onnxruntime_providers_cuda_plugin
+      CUDA_ARCHITECTURES "${_plugin_xqa_cuda_architectures}"
+      NVCC_THREADS "${onnxruntime_plugin_nvcc_threads}"
+      COMPILE_OPTIONS ${_cuda_plugin_shared_compile_options}
+      SOURCES ${_cuda_plugin_xqa_srcs})
+  else()
+    target_sources(onnxruntime_providers_cuda_plugin PRIVATE ${_cuda_plugin_xqa_srcs})
+  endif()
+endif()
+
 if(NOT onnxruntime_DISABLE_CONTRIB_OPS)
   # SM90 TMA warp-specialized files use SM90-specific collective operations.
   # Also includes fpA_intB SM90 launchers (guarded by #ifndef EXCLUDE_SM_90).
diff --git a/cmake/onnxruntime_unittests.cmake b/cmake/onnxruntime_unittests.cmake
index 1974b5d601f46..40669a6681aaf 100644
--- a/cmake/onnxruntime_unittests.cmake
+++ b/cmake/onnxruntime_unittests.cmake
@@ -1036,6 +1036,9 @@ if (onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS AND NOT onnxruntime_BUILD_CUDA_EP_
   if(TARGET onnxruntime_providers_cuda_flash_attention)
     target_link_libraries(onnxruntime_providers_cuda_ut PRIVATE onnxruntime_providers_cuda_flash_attention)
   endif()
+  if(TARGET onnxruntime_providers_cuda_xqa)
+    target_link_libraries(onnxruntime_providers_cuda_ut PRIVATE onnxruntime_providers_cuda_xqa)
+  endif()
   if(TARGET onnxruntime_providers_cuda_llm)
     target_link_libraries(onnxruntime_providers_cuda_ut PRIVATE onnxruntime_providers_cuda_llm)
   endif()