Create seperate cmake target for sampler-related cuda functions

backends/cuda/CMakeLists.txt

@@ -110,8 +110,7 @@ set(_aoti_cuda_shim_sources runtime/cuda_allocator.cpp runtime/shims/memory.cpp
 # Only build CUDA shims when CUDA language/toolchain is available.
 if(CMAKE_CUDA_COMPILER)
   list(APPEND _aoti_cuda_shim_sources runtime/shims/int4mm.cu
-       runtime/shims/int4_plain_mm.cu runtime/shims/sort.cu
-       runtime/shims/rand.cu
+       runtime/shims/sort.cu runtime/shims/int4_plain_mm.cu
   )
 endif()
 
@@ -153,8 +152,7 @@ endif()
 # retention.
 if(_cuda_is_msvc_toolchain)
   target_link_libraries(
-    aoti_cuda_shims PRIVATE cuda_platform CUDA::cudart CUDA::curand
-                            ${CMAKE_DL_LIBS}
+    aoti_cuda_shims PRIVATE cuda_platform CUDA::cudart ${CMAKE_DL_LIBS}
   )
   # Link object library directly so symbols are pulled exactly once while
   # avoiding duplicate static/object inclusion and interface leakage.
@@ -164,7 +162,7 @@ else()
     aoti_cuda_shims
     PRIVATE cuda_platform
     PUBLIC -Wl,--whole-archive aoti_common_shims_slim -Wl,--no-whole-archive
-           CUDA::cudart CUDA::curand ${CMAKE_DL_LIBS}
+           CUDA::cudart ${CMAKE_DL_LIBS}
   )
 endif()
 
@@ -178,6 +176,83 @@ install(
   DESTINATION lib
 )
 
+# CUDA-specific AOTI sampler shim symbols (rand/randint via curand). Split out
+# of aoti_cuda_shims so the curand fatbin (~3.5MB precalc tables + Philox
+# kernels per arch) and the CUDA::curand dependency are only paid by the small
+# set of consumers that actually use them (e.g. qwen3_5_moe). Other CUDA
+# examples (voxtral, parakeet, whisper, dinov2, ...) link only aoti_cuda_shims
+# and stay small.
+if(CMAKE_CUDA_COMPILER)
+  add_library(aoti_cuda_sampler_shims SHARED runtime/shims/rand.cu)
+
+  # Match aoti_cuda_shims preprocessor defines for symbol export.
+  target_compile_definitions(aoti_cuda_sampler_shims PRIVATE CUDA_AVAILABLE=1)
+  if(WIN32)
+    target_compile_definitions(
+      aoti_cuda_sampler_shims PRIVATE EXPORT_AOTI_FUNCTIONS
+    )
+    if(_cuda_is_windows_msvc)
+      set_target_properties(
+        aoti_cuda_sampler_shims PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS OFF
+      )
+    endif()
+  endif()
+
+  target_include_directories(
+    aoti_cuda_sampler_shims
+    PUBLIC ${CUDAToolkit_INCLUDE_DIRS} $<BUILD_INTERFACE:${EXECUTORCH_ROOT}>
+           $<INSTALL_INTERFACE:include>
+  )
+
+  target_compile_options(
+    aoti_cuda_sampler_shims
+    PUBLIC "$<$<COMPILE_LANGUAGE:CXX>:${_cuda_cxx_compile_options}>"
+  )
+
+  if(_cuda_export_dynamic_option)
+    target_link_options(
+      aoti_cuda_sampler_shims PUBLIC ${_cuda_export_dynamic_option}
+    )
+  endif()
+
+  # rand.cu calls into slim helpers (empty_strided, getCurrentCUDAStream,
+  # SlimTensor) which are linked into aoti_cuda_shims. Depend on that target so
+  # we resolve those symbols from the already-loaded shims library instead of
+  # duplicating slim's static archive into both DLLs.
+  #
+  # Also link `slimtensor` (INTERFACE / header-only) directly so the c10 include
+  # root (runtime/core/portable_type/c10) is on this target's compile command.
+  # aoti_cuda_shims links aoti_common_shims_slim PUBLIC on non-MSVC (so includes
+  # propagate transitively on Linux) but only PRIVATELY via the *_obj OBJECT lib
+  # on MSVC, which does NOT forward the slimtensor INTERFACE include dirs.
+  # Linking slimtensor here makes the include path explicit on both toolchains
+  # and keeps Windows MSVC happy without changing aoti_cuda_shims' propagation
+  # semantics.
+  if(_cuda_is_msvc_toolchain)
+    target_link_libraries(
+      aoti_cuda_sampler_shims
+      PRIVATE cuda_platform CUDA::cudart CUDA::curand ${CMAKE_DL_LIBS}
+              aoti_cuda_shims slimtensor
+    )
+  else()
+    target_link_libraries(
+      aoti_cuda_sampler_shims
+      PRIVATE cuda_platform slimtensor
+      PUBLIC CUDA::cudart CUDA::curand ${CMAKE_DL_LIBS} aoti_cuda_shims
+    )
+  endif()
+
+  if(NOT _cuda_is_msvc_toolchain)
+    executorch_target_link_options_shared_lib(aoti_cuda_sampler_shims)
+  endif()
+
+  install(
+    TARGETS aoti_cuda_sampler_shims
+    EXPORT ExecuTorchTargets
+    DESTINATION lib
+  )
+endif()
+
 # CUDA backend implementation
 set(_aoti_cuda_backend_sources runtime/cuda_backend.cpp)
 if(_cuda_is_msvc_toolchain)

backends/cuda/runtime/shims/tests/CMakeLists.txt

@@ -90,3 +90,7 @@ foreach(test_name ${CUDA_KERNEL_TESTS})
 
   add_test(NAME ${test_name} COMMAND ${test_name})
 endforeach()
+
+# rand symbols live in the separate aoti_cuda_sampler_shims DLL to keep the
+# curand-induced binary-size cost out of aoti_cuda_shims.
+target_link_libraries(test_aoti_torch_cuda_rand PRIVATE aoti_cuda_sampler_shims)

examples/models/qwen3_5_moe/CMakeLists.txt

@@ -48,8 +48,9 @@ if(EXECUTORCH_BUILD_METAL)
   executorch_target_link_options_shared_lib(metal_backend)
 elseif(EXECUTORCH_BUILD_CUDA)
   find_package(CUDAToolkit REQUIRED)
-  list(APPEND link_libraries aoti_cuda_backend)
+  list(APPEND link_libraries aoti_cuda_backend aoti_cuda_sampler_shims)
   executorch_target_link_options_shared_lib(aoti_cuda_backend)
+  executorch_target_link_options_shared_lib(aoti_cuda_sampler_shims)
   add_compile_definitions(EXECUTORCH_BUILD_CUDA)
 else()
   message(