Implement Realm reductions for CUDA

.flake/pkgs/realm.nix

@@ -13,13 +13,13 @@ in
 
 stdenv.mkDerivation rec {
   pname = "realm";
-  version = "2026-07-01";
+  version = "2026-08-28";
 
   src = fetchFromGitHub {
     owner = "StanfordLegion";
     repo = "realm";
-    rev = "dba83452d4a211e3a0c0e5d4849c26ce7b56bba3";
-    sha256 = "sha256-dLGmUgupDX0dDErZbuP5ipj2Cwf+rOhVGC/+angRyBY=";
+    rev = "15b3e9b68b65d698d237766d37bc36522b8ecf0b";
+    sha256 = "sha256-HXNRKzOIUok8FAuhJF1sOW0+Hv7rRfRDoQNt+wAvOnc=";
   };
 
   nativeBuildInputs = [

CMakeLists.txt

@@ -51,15 +51,17 @@ option(FF_BUILD_ARG_PARSER "build command line argument parser" OFF)
 option(FF_BUILD_BIN_EXPORT_MODEL_ARCH "build export-model-arch utility" ON)
 option(FF_BUILD_BIN_RUN_MODEL "build run-model binary" ON)
 
-include(cuda)
-include(cudnn)
-include(nccl)
 if (FF_USE_CODE_COVERAGE)
   include(CodeCoverage)
   append_coverage_compiler_flags()
 endif()
 # set_property(CACHE FF_GPU_BACKEND PROPERTY STRINGS ${FF_GPU_BACKENDS})
 
+# Load after CodeCoverage because CodeCoverage blows up on NVCC compiler
+include(cuda)
+include(cudnn)
+include(nccl)
+
 include(json)
 include(expected)
 include(spdlog)

cmake/cuda.cmake

@@ -1,5 +1,6 @@
 include(aliasing)
 
+enable_language(CUDA)
 find_package(CUDAToolkit 11.7 REQUIRED)
 
 alias_library(deps::cublas CUDA::cublas)

cmake/flexflow-utils.cmake

@@ -56,13 +56,22 @@ function(ff_set_cxx_properties target)
   )
 endfunction()
 
+function(ff_set_cuda_properties target)
+  set_target_properties(${target}
+    PROPERTIES
+      CUDA_STANDARD 17
+      CUDA_STANDARD_REQUIRED YES
+  )
+endfunction()
+
 function(ff_add_library)
   ff_parse_args(
     PREFIX
       FF_LIBRARY
     ARGS
       NAME
     VARIADIC_ARGS
+      LANGUAGES
       SRC_PATTERNS
       PUBLIC_INCLUDE
       PRIVATE_INCLUDE
@@ -72,7 +81,11 @@ function(ff_add_library)
       ${ARGN}
   )
 
-  project(${FF_LIBRARY_NAME})
+  if(NOT FF_LIBRARY_LANGUAGES)
+    project(${FF_LIBRARY_NAME})
+  else()
+    project(${FF_LIBRARY_NAME} LANGUAGES ${FF_LIBRARY_LANGUAGES})
+  endif()
   file(GLOB_RECURSE SRC
        CONFIGURE_DEPENDS
        LIST_DIRECTORIES False
@@ -99,6 +112,9 @@ function(ff_add_library)
   )
   define_ff_vars(${FF_LIBRARY_NAME})
   ff_set_cxx_properties(${FF_LIBRARY_NAME})
+  if(CUDA IN_LIST FF_LIBRARY_LANGUAGES)
+    ff_set_cuda_properties(${FF_LIBRARY_NAME})
+  endif()
 endfunction()
 
 function(ff_add_test_executable)

lib/realm-execution/CMakeLists.txt

@@ -1,8 +1,12 @@
 ff_add_library(
   NAME
     realm-execution
+  LANGUAGES
+    CXX
+    CUDA
   SRC_PATTERNS
     src/*.cc
+    src/*.cu
   PUBLIC_INCLUDE
     include/
   PRIVATE_INCLUDE