Tidy up patches

recipe/meta.yaml

@@ -18,9 +18,7 @@ source:
     - patches/0001-Force-endian-flag-in-cross-compilation-mode.patch  # [arm64 or aarch64 or ppc64le]
     - patches/0002-Enable-latest-libcxx-on-MacOS.patch  # [osx]
     - patches/0003-Use-mingw-w64-path.patch
-    - patches/0004-Undo-dmlc-xgboost-9436.patch
-    - patches/0005-CMake-Explicitly-link-with-CCCL-standalone-or-CTK.patch
-    - patches/0006-Partial-fix-for-CTK-12.5-10574.patch
+    - patches/0004-Backport-fixes.patch
 
 build:
   number: {{ build_number }}

recipe/patches/0004-Backport-fixes.patch

@@ -0,0 +1,143 @@
+From 7782edc8e004bd3fe10879c52753d6a6de35947d Mon Sep 17 00:00:00 2001
+From: Hyunsu Cho <phcho@nvidia.com>
+Date: Fri, 28 Jun 2024 17:38:34 -0700
+Subject: [PATCH] Backport fixes
+
+* Undo dmlc/xgboost#9436
+* Explicitly link with CCCL (standalone or CTK) (#10624, #10633)
+* Partial fix for CTK 12.5 (#10574)
+---
+ CMakeLists.txt                  | 27 ++++++++++++++++++---------
+ cmake/Utils.cmake               |  7 +++----
+ src/tree/updater_gpu_common.cuh | 21 ++++++++-------------
+ src/tree/updater_gpu_hist.cu    |  1 +
+ 4 files changed, 30 insertions(+), 26 deletions(-)
+
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index c4ca82937..aa078405f 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -224,6 +224,24 @@ if(USE_CUDA)
+   add_subdirectory(${PROJECT_SOURCE_DIR}/gputreeshap)
+
+   find_package(CUDAToolkit REQUIRED)
++  find_package(CCCL CONFIG)
++  if(NOT CCCL_FOUND)
++    message(STATUS "Standalone CCCL not found. Attempting to use CCCL from CUDA Toolkit...")
++    find_package(CCCL CONFIG
++      HINTS ${CUDAToolkit_LIBRARY_DIR}/cmake)
++    if(NOT CCCL_FOUND)
++      message(STATUS "Could not locate CCCL from CUDA Toolkit. Using Thrust and CUB from CUDA Toolkit...")
++      find_package(libcudacxx CONFIG REQUIRED
++        HINTS ${CUDAToolkit_LIBRARY_DIR}/cmake)
++      find_package(CUB CONFIG REQUIRED
++        HINTS ${CUDAToolkit_LIBRARY_DIR}/cmake)
++      find_package(Thrust CONFIG REQUIRED
++        HINTS ${CUDAToolkit_LIBRARY_DIR}/cmake)
++      thrust_create_target(Thrust HOST CPP DEVICE CUDA)
++      add_library(CCCL::CCCL INTERFACE IMPORTED GLOBAL)
++      target_link_libraries(CCCL::CCCL INTERFACE libcudacxx::libcudacxx CUB::CUB Thrust)
++    endif()
++  endif()
+ endif()
+
+ if(FORCE_COLORED_OUTPUT AND (CMAKE_GENERATOR STREQUAL "Ninja") AND
+@@ -301,15 +319,6 @@ add_subdirectory(${xgboost_SOURCE_DIR}/plugin)
+
+ if(PLUGIN_RMM)
+   find_package(rmm REQUIRED)
+-
+-  # Patch the rmm targets so they reference the static cudart
+-  # Remove this patch once RMM stops specifying cudart requirement
+-  # (since RMM is a header-only library, it should not specify cudart in its CMake config)
+-  get_target_property(rmm_link_libs rmm::rmm INTERFACE_LINK_LIBRARIES)
+-  list(REMOVE_ITEM rmm_link_libs CUDA::cudart)
+-  list(APPEND rmm_link_libs CUDA::cudart_static)
+-  set_target_properties(rmm::rmm PROPERTIES INTERFACE_LINK_LIBRARIES "${rmm_link_libs}")
+-  get_target_property(rmm_link_libs rmm::rmm INTERFACE_LINK_LIBRARIES)
+ endif()
+
+ if(PLUGIN_SYCL)
+diff --git a/cmake/Utils.cmake b/cmake/Utils.cmake
+index d555f5edf..c1f19fdbb 100644
+--- a/cmake/Utils.cmake
++++ b/cmake/Utils.cmake
+@@ -108,12 +108,12 @@ function(xgboost_set_cuda_flags target)
+     target_compile_definitions(${target} PRIVATE -DXGBOOST_USE_NVTX=1)
+   endif()
+
++  target_link_libraries(${target}
++    PRIVATE CCCL::CCCL CUDA::cudart_static)
+   target_compile_definitions(${target} PRIVATE -DXGBOOST_USE_CUDA=1)
+   target_include_directories(
+     ${target} PRIVATE
+-    ${xgboost_SOURCE_DIR}/gputreeshap
+-    ${xgboost_SOURCE_DIR}/rabit/include
+-    ${CUDAToolkit_INCLUDE_DIRS})
++    ${xgboost_SOURCE_DIR}/gputreeshap)
+
+   if(MSVC)
+     target_compile_options(${target} PRIVATE
+@@ -240,7 +240,6 @@ macro(xgboost_target_link_libraries target)
+
+   if(USE_CUDA)
+     xgboost_set_cuda_flags(${target})
+-    target_link_libraries(${target} PUBLIC CUDA::cudart_static)
+   endif()
+
+   if(PLUGIN_RMM)
+diff --git a/src/tree/updater_gpu_common.cuh b/src/tree/updater_gpu_common.cuh
+index 1c3e6a552..5d999d6d6 100644
+--- a/src/tree/updater_gpu_common.cuh
++++ b/src/tree/updater_gpu_common.cuh
+@@ -1,18 +1,13 @@
+-/*!
+- * Copyright 2017-2019 XGBoost contributors
++/**
++ * Copyright 2017-2024, XGBoost contributors
+  */
+ #pragma once
+-#include <thrust/random.h>
+-#include <cstdio>
+-#include <cub/cub.cuh>
+-#include <stdexcept>
+-#include <string>
+-#include <vector>
+-#include "../common/categorical.h"
+-#include "../common/device_helpers.cuh"
+-#include "../common/random.h"
++#include <limits>   // for numeric_limits
++#include <ostream>  // for ostream
++
+ #include "gpu_hist/histogram.cuh"
+ #include "param.h"
++#include "xgboost/base.h"
+
+ namespace xgboost::tree {
+ struct GPUTrainingParam {
+@@ -54,8 +49,8 @@ enum DefaultDirection {
+ };
+
+ struct DeviceSplitCandidate {
+-  float loss_chg {-FLT_MAX};
+-  DefaultDirection dir {kLeftDir};
++  float loss_chg{-std::numeric_limits<float>::max()};
++  DefaultDirection dir{kLeftDir};
+   int findex {-1};
+   float fvalue {0};
+   // categorical split, either it's the split category for OHE or the threshold for partition-based
+diff --git a/src/tree/updater_gpu_hist.cu b/src/tree/updater_gpu_hist.cu
+index 958fa0331..e126aeb31 100644
+--- a/src/tree/updater_gpu_hist.cu
++++ b/src/tree/updater_gpu_hist.cu
+@@ -19,6 +19,7 @@
+ #include "../common/cuda_context.cuh"  // CUDAContext
+ #include "../common/device_helpers.cuh"
+ #include "../common/hist_util.h"
++#include "../common/random.h"  // for ColumnSampler, GlobalRandom
+ #include "../common/timer.h"
+ #include "../data/ellpack_page.cuh"
+ #include "../data/ellpack_page.h"
+-- 
+2.45.2
+