pytorch · jackbondpreston-arm · Oct 10, 2025 · Oct 10, 2025 · Oct 10, 2025 · Oct 10, 2025
diff --git a/.github/workflows/build-linux.yml b/.github/workflows/build-linux.yml
@@ -23,6 +23,8 @@ jobs:
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
+        with:
+          submodules: true
       - name: Setup environment
         run: |
           sudo apt-get install -y libssl-dev

diff --git a/.github/workflows/build-windows.yml b/.github/workflows/build-windows.yml
@@ -24,6 +24,8 @@ jobs:
         uses: microsoft/setup-msbuild@v2
       - name: Checkout code
         uses: actions/checkout@v4
+        with:
+          submodules: true
       - name: Set up Python 3.10
         uses: actions/setup-python@v5
         with:

diff --git a/.github/workflows/super-linter.yml b/.github/workflows/super-linter.yml
@@ -24,6 +24,7 @@ jobs:
           # Full git history is needed to get a proper list of changed files
           # within `super-linter`
           fetch-depth: 0
+          submodules: true
       - name: Super-linter
         uses: super-linter/super-linter@v7.3.0 # x-release-please-version
         env:

diff --git a/.gitmodules b/.gitmodules
@@ -0,0 +1,3 @@
+[submodule "third_party/spdlog"]
+	path = third_party/spdlog
+	url = https://github.com/gabime/spdlog.git
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -98,6 +98,9 @@ endif()
 set(GLOO_INSTALL ON CACHE BOOL "")
 mark_as_advanced(GLOO_INSTALL)
 
+# Compile definitions for Gloo
+set(GLOO_COMPILE_DEFS)
+
 # Build shared or static libraries (override from parent project)
 if(BUILD_SHARED_LIBS)
   set(GLOO_STATIC_OR_SHARED SHARED CACHE STRING "")

diff --git a/README.md b/README.md
@@ -62,6 +62,23 @@ Optional dependencies are:
 
 Please refer to [docs/](docs/) for detailed documentation.
 
+## Cloning
+
+Gloo makes use of [Git submodules](https://git-scm.com/book/en/v2/Git-Tools-Submodules)
+for third-party build-time dependencies. When cloning, ensure that you
+initialize and sync submodules:
+
+``` shell
+git clone --recurse-submodules https://github.com/pytorch/gloo.git # or use SSH remote
+```
+
+Or, if you already cloned without submodules, run the following in the existing
+local repository:
+
+``` shell
+git submodule update --init --recursive
+```
+
 ## Building
 
 You can build Gloo using CMake.

diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
@@ -2,6 +2,9 @@ set(gloo_DEPENDENCY_LIBS "")
 set(gloo_cuda_DEPENDENCY_LIBS "")
 set(gloo_hip_DEPENDENCY_LIBS "")
 
+# Dependency libraries for all Gloo targets (e.g. tests, benchmarks, etc.)
+set(gloo_ALL_TARGETS_DEPENDENCY_LIBS "")
+
 # Configure path to modules (for find_package)
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${PROJECT_SOURCE_DIR}/cmake/Modules/")
 
@@ -104,7 +107,7 @@ if(USE_MPI)
     message(STATUS "MPI libraries: " ${MPI_CXX_LIBRARIES})
     include_directories(SYSTEM ${MPI_CXX_INCLUDE_PATH})
     list(APPEND gloo_DEPENDENCY_LIBS ${MPI_CXX_LIBRARIES})
-    add_definitions(-DGLOO_USE_MPI=1)
+    list(APPEND GLOO_COMPILE_DEFS "GLOO_USE_MPI=1")
   else()
     message(WARNING "Not compiling with MPI support. Suppress this warning with -DUSE_MPI=OFF.")
     set(USE_MPI OFF)
@@ -199,3 +202,19 @@ if(BUILD_TEST)
     target_link_libraries(gtest INTERFACE ${GTEST_LIBRARIES})
   endif()
 endif()
+
+set(SPDLOG_BUILD_SHARED OFF CACHE BOOL "" FORCE)
+set(SPDLOG_BUILD_PIC ON CACHE BOOL "" FORCE)
+set(SPDLOG_FMT_EXTERNAL OFF CACHE BOOL "" FORCE)
+
+set(_save_BUILD_SHARED_LIBS ${BUILD_SHARED_LIBS})
+set(BUILD_SHARED_LIBS OFF)
+add_subdirectory(third_party/spdlog EXCLUDE_FROM_ALL)
+set(BUILD_SHARED_LIBS ${_save_BUILD_SHARED_LIBS})
+
+if(GLOO_STATIC_OR_SHARED STREQUAL "STATIC")
+  list(APPEND gloo_ALL_TARGETS_DEPENDENCY_LIBS $<BUILD_INTERFACE:spdlog::spdlog_header_only>)
+else()
+  list(APPEND gloo_ALL_TARGETS_DEPENDENCY_LIBS spdlog::spdlog)
+  list(APPEND GLOO_COMPILE_DEFS SPDLOG_COMPILED_LIB)
+endif()
diff --git a/docs/errors.md b/docs/errors.md
@@ -18,4 +18,4 @@ extend `::gloo::Exception`.
 ## Assertions
 Gloo asserts unexpected errors and logical invariants instead of expecting
 callers to handle them. `GLOO_ENFORCE` macros are defined in
-[`logging.h`](../gloo/common/logging.h)
+[`enforce.h`](../gloo/common/enforce.h)
diff --git a/gloo/CMakeLists.txt b/gloo/CMakeLists.txt
@@ -147,10 +147,12 @@ if(USE_CUDA)
     cuda_add_library(gloo_cuda ${GLOO_CUDA_SRCS} ${GLOO_STATIC_OR_SHARED})
   endif()
   target_link_libraries(gloo_cuda gloo ${gloo_cuda_DEPENDENCY_LIBS})
+  set_property(TARGET gloo_cuda APPEND PROPERTY LINK_LIBRARIES ${gloo_ALL_TARGETS_DEPENDENCY_LIBS})
 endif()
 if(USE_ROCM)
   gloo_hip_add_library(gloo_hip ${GLOO_HIP_SRCS})
   target_link_libraries(gloo_hip gloo)
+  set_property(TARGET gloo_hip APPEND PROPERTY LINK_LIBRARIES ${gloo_ALL_TARGETS_DEPENDENCY_LIBS})
 endif()
 if(USE_LIBUV)
   target_link_libraries(gloo PRIVATE uv_a)
@@ -165,7 +167,8 @@ elseif(USE_TCP_OPENSSL_LOAD)
   target_link_libraries(gloo PRIVATE dl)
 endif()
 
-target_link_libraries(gloo PRIVATE ${gloo_DEPENDENCY_LIBS})
+target_link_libraries(gloo PRIVATE ${gloo_DEPENDENCY_LIBS} ${gloo_ALL_TARGETS_DEPENDENCY_LIBS})
+target_compile_definitions(gloo PRIVATE ${GLOO_COMPILE_DEFS})
 
 # Add Interface include directories that are relocatable.
 target_include_directories(gloo INTERFACE $<INSTALL_INTERFACE:include>)

diff --git a/gloo/algorithm.cc b/gloo/algorithm.cc
@@ -8,7 +8,7 @@
 
 #include "gloo/algorithm.h"
 
-#include "gloo/common/logging.h"
+#include "gloo/common/enforce.h"
 
 namespace gloo {
 

diff --git a/gloo/allgather.cc b/gloo/allgather.cc
@@ -11,7 +11,7 @@
 #include <array>
 #include <cstring>
 
-#include "gloo/common/logging.h"
+#include "gloo/common/enforce.h"
 #include "gloo/types.h"
 
 namespace gloo {

diff --git a/gloo/allgatherv.cc b/gloo/allgatherv.cc
@@ -11,7 +11,7 @@
 #include <cstring>
 #include <numeric>
 
-#include "gloo/common/logging.h"
+#include "gloo/common/enforce.h"
 #include "gloo/types.h"
 
 namespace gloo {

diff --git a/gloo/allreduce.cc b/gloo/allreduce.cc
@@ -12,7 +12,7 @@
 #include <array>
 #include <cstring>
 
-#include "gloo/common/logging.h"
+#include "gloo/common/enforce.h"
 #include "gloo/math.h"
 #include "gloo/types.h"
 

diff --git a/gloo/allreduce_bcube.cc b/gloo/allreduce_bcube.cc
@@ -0,0 +1,74 @@
+#include "gloo/allreduce_bcube.h"
+
+#include <string_view>
+
+#include "gloo/common/log.h"
+
+namespace gloo {
+
+template <typename T>
+void AllreduceBcube<T>::printElems(T* p, int count, int start) {
+  /* Early return if log level is not high enough, to prevent expensive code
+   * running. */
+  if (!spdlog::should_log(spdlog::level::trace))
+    return;
+
+  const std::size_t alignedStart = (start / wordsPerLine) * wordsPerLine;
+  fmt::memory_buffer line{};
+
+  /* Logs/flushes the line buffer - starting a new line */
+  auto printLine = [&]() {
+    if (!line.size())
+      return;
+    std::string_view sv{line.data(), line.size()};
+    GLOO_TRACE("{}", sv);
+    line.clear();
+  };
+
+  for (std::size_t x = alignedStart; x < start + count; ++x) {
+    if (x % wordsPerLine == 0) {
+      if (x != alignedStart)
+        printLine();
+      fmt::format_to(
+          std::back_inserter(line), "{} {:05}: ", fmt::ptr(&p[x]), x);
+    } else if (x % wordsPerSection == 0) {
+      fmt::format_to(std::back_inserter(line), "- ");
+    }
+
+    if (x < start)
+      fmt::format_to(std::back_inserter(line), "..... ");
+    else
+      fmt::format_to(std::back_inserter(line), "{:05} ", p[x]);
+  }
+  printLine();
+}
+
+template <typename T>
+void AllreduceBcube<T>::printStageBuffer(const std::string& msg) {
+  if (printCheck(myRank_)) {
+    GLOO_TRACE("rank ({}) {}:", myRank_, msg);
+    printElems(&ptrs_[0][0], totalNumElems_);
+  }
+}
+
+template <typename T>
+void AllreduceBcube<T>::printStepBuffer(
+    const std::string& stage,
+    int step,
+    int srcRank,
+    int destRank,
+    T* p,
+    int count,
+    int start) {
+  if (printCheck(myRank_)) {
+    GLOO_TRACE(
+        "{}: step ({}) srcRank ({}) -> destRank ({})",
+        stage,
+        step,
+        srcRank,
+        destRank);
+    printElems(p, count, start);
+  }
+}
+
+} // namespace gloo
diff --git a/gloo/allreduce_bcube.h b/gloo/allreduce_bcube.h
@@ -13,15 +13,12 @@
 #include <string.h>
 #include <algorithm>
 #include <cstring>
-#include <iomanip>
-#include <iostream>
 #include <memory>
 #include <string>
 #include <unordered_map>
 #include <vector>
 
 #include "gloo/algorithm.h"
-#include "gloo/common/error.h"
 #include "gloo/context.h"
 
 /**
@@ -537,48 +534,18 @@ class AllreduceBcube : public Algorithm {
   static bool printCheck(int /*rank*/) {
     return false;
   }
-  /**
-   * Prints a break given the offset of an element about to be printed
-   * @param p Pointer to the elements
-   * @param x The current offset to the pointer to words
-   */
-  static void printBreak(T* p, int x) {
-    if (0 == x % wordsPerLine) {
-      std::cout << std::endl
-                << &p[x] << " " << std::setfill('0') << std::setw(5) << x
-                << ": ";
-    } else if (0 == x % wordsPerSection) {
-      std::cout << "- ";
-    }
-  }
   /**
    * Pretty prints a list of elements
    * @param p Pointer to the elements
    * @param count The number of elements to be printed
    * @param start The offset from which to print
    */
-  static void printElems(T* p, int count, int start = 0) {
-    auto alignedStart = (start / wordsPerLine) * wordsPerLine;
-    for (int x = alignedStart; x < start + count; ++x) {
-      printBreak(p, x);
-      if (x < start) {
-        std::cout << "..... ";
-      } else {
-        std::cout << std::setfill('0') << std::setw(5) << p[x] << " ";
-      }
-    }
-  }
+  static void printElems(T* p, int count, int start);
   /**
    * Prints contents in the ptrs array at a particular stage
    * @param msg Custom message to be printed
    */
-  void printStageBuffer(const std::string& msg) {
-    if (printCheck(myRank_)) {
-      std::cout << "rank (" << myRank_ << ") " << msg << ": ";
-      printElems(&ptrs_[0][0], totalNumElems_);
-      std::cout << std::endl;
-    }
-  }
+  void printStageBuffer(const std::string& msg);
 
   /**
    * Prints specified buffer during a step
@@ -596,14 +563,7 @@ class AllreduceBcube : public Algorithm {
       int destRank,
       T* p,
       int count,
-      int start = 0) {
-    if (printCheck(myRank_)) {
-      std::cout << stage << ": step (" << step << ") " << "srcRank (" << srcRank
-                << ") -> " << "destRank (" << destRank << "): ";
-      printElems(p, count, start);
-      std::cout << std::endl;
-    }
-  }
+      int start = 0);
   /**
    * Get all the peers of node with specified rank
    * @param rank Rank of the node for which peers are needed

diff --git a/gloo/alltoall.cc b/gloo/alltoall.cc
@@ -10,7 +10,7 @@
 
 #include <cstring>
 
-#include "gloo/common/logging.h"
+#include "gloo/common/enforce.h"
 #include "gloo/types.h"
 
 namespace gloo {

diff --git a/gloo/alltoall.h b/gloo/alltoall.h
@@ -8,7 +8,7 @@
 
 #pragma once
 
-#include "gloo/common/logging.h"
+#include "gloo/common/enforce.h"
 #include "gloo/context.h"
 #include "gloo/transport/unbound_buffer.h"
 

diff --git a/gloo/alltoallv.cc b/gloo/alltoallv.cc
@@ -11,7 +11,7 @@
 #include <cstring>
 #include <numeric>
 
-#include "gloo/common/logging.h"
+#include "gloo/common/enforce.h"
 #include "gloo/types.h"
 
 namespace gloo {

diff --git a/gloo/alltoallv.h b/gloo/alltoallv.h
@@ -8,7 +8,7 @@
 
 #pragma once
 
-#include "gloo/common/logging.h"
+#include "gloo/common/enforce.h"
 #include "gloo/context.h"
 #include "gloo/transport/unbound_buffer.h"
 

diff --git a/gloo/benchmark/CMakeLists.txt b/gloo/benchmark/CMakeLists.txt
@@ -5,7 +5,7 @@ set(GLOO_BENCHMARK_SRCS
   )
 
 add_executable(benchmark ${GLOO_BENCHMARK_SRCS})
-target_link_libraries(benchmark gloo)
+target_link_libraries(benchmark gloo ${gloo_ALL_TARGETS_DEPENDENCY_LIBS})
 
 if(GLOO_INSTALL)
   install(TARGETS benchmark DESTINATION ${CMAKE_INSTALL_PREFIX}/bin)

diff --git a/gloo/benchmark/cuda_main.cc b/gloo/benchmark/cuda_main.cc
@@ -11,7 +11,7 @@
 
 #include "gloo/benchmark/benchmark.h"
 #include "gloo/benchmark/runner.h"
-#include "gloo/common/logging.h"
+#include "gloo/common/enforce.h"
 #include "gloo/cuda_allreduce_bcube.h"
 #include "gloo/cuda_allreduce_halving_doubling.h"
 #include "gloo/cuda_allreduce_halving_doubling_pipelined.h"

diff --git a/gloo/benchmark/main.cc b/gloo/benchmark/main.cc
@@ -26,7 +26,7 @@
 #include "gloo/broadcast_one_to_all.h"
 #include "gloo/common/aligned_allocator.h"
 #include "gloo/common/common.h"
-#include "gloo/common/logging.h"
+#include "gloo/common/enforce.h"
 #include "gloo/context.h"
 #include "gloo/pairwise_exchange.h"
 #include "gloo/reduce.h"

diff --git a/gloo/benchmark/runner.cc b/gloo/benchmark/runner.cc
@@ -15,7 +15,7 @@
 #include "gloo/barrier_all_to_one.h"
 #include "gloo/broadcast_one_to_all.h"
 #include "gloo/common/common.h"
-#include "gloo/common/logging.h"
+#include "gloo/common/enforce.h"
 #include "gloo/rendezvous/context.h"
 #include "gloo/rendezvous/file_store.h"
 #include "gloo/rendezvous/prefix_store.h"