Graph parallel: the next generation (#1080)

* WIP: absorb adding input into std_attn and std_ffn

* WIP: NCCL infra

* WIP: add reduce and fake_cpy ops

* WIP

* WIP: graph appears to work, layer is broken

* WIP: Qwen3-MoE works with graph, layer still broken

* WIP: GLM-4.5 graph works

* WIP: fix sm layer (dense)

* WIP: fix sm layer (MoE)

* WIP: fast PP with bespoke 4-GPU NCCL

I guess, I'm not using NCCL the right way as PP is very
low with a single communicator group for 3 or more GPUs.
But if I create 4 communicator groups for pairs of GPUs
(0,1, 2,3, 0,2, 1,3) and use that, PP is fast: I'm hitting
1500 t/s for L3-70B on the 4x3090 system, which is
~20% better than the previous sm graph without NCCL.
But that cannot be the solution (I cannot be creating pairwise
communicators and associated logic for every possible number of GPUs).

* WIP: Cohere2

* Explicitely set device

* Bespoke 3-GPU case

* WIP

* Do not repeat get_rows multiple times

* Fix 3 GPUs

* OK, let's leave it in

* Implement the reduce op without NCCL available

* Be able to build without NCCL

cmake -DGGML_NCCL=OFF disables it

* Make --max-gpu work again

* Slightly better for 4 GPUs without NCCL

* Cleanup

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-12-24 08:31:48 +01:00
committed by GitHub
parent 2a633c4357
commit 0d7eb34185
12 changed files with 870 additions and 256 deletions

90
cmake/FindNCCL.cmake Normal file
View File

@@ -0,0 +1,90 @@
# Find the nccl libraries
#
# The following variables are optionally searched for defaults
# NCCL_ROOT: Base directory where all NCCL components are found
# NCCL_INCLUDE_DIR: Directory where NCCL header is found
# NCCL_LIB_DIR: Directory where NCCL library is found
#
# The following are set after configuration is done:
# NCCL_FOUND
# NCCL_INCLUDE_DIRS
# NCCL_LIBRARIES
#
# The path hints include CUDA_TOOLKIT_ROOT_DIR seeing as some folks
# install NCCL in the same location as the CUDA toolkit.
# See https://github.com/caffe2/caffe2/issues/1601
set(NCCL_INCLUDE_DIR $ENV{NCCL_INCLUDE_DIR} CACHE PATH "Folder contains NVIDIA NCCL headers")
set(NCCL_LIB_DIR $ENV{NCCL_LIB_DIR} CACHE PATH "Folder contains NVIDIA NCCL libraries")
set(NCCL_VERSION $ENV{NCCL_VERSION} CACHE STRING "Version of NCCL to build with")
if ($ENV{NCCL_ROOT_DIR})
message(WARNING "NCCL_ROOT_DIR is deprecated. Please set NCCL_ROOT instead.")
endif()
list(APPEND NCCL_ROOT $ENV{NCCL_ROOT_DIR} ${CUDA_TOOLKIT_ROOT_DIR})
# Compatible layer for CMake <3.12. NCCL_ROOT will be accounted in for searching paths and libraries for CMake >=3.12.
list(APPEND CMAKE_PREFIX_PATH ${NCCL_ROOT})
find_path(NCCL_INCLUDE_DIRS
NAMES nccl.h
HINTS ${NCCL_INCLUDE_DIR})
if (USE_STATIC_NCCL)
MESSAGE(STATUS "USE_STATIC_NCCL is set. Linking with static NCCL library.")
SET(NCCL_LIBNAME "nccl_static")
if (NCCL_VERSION) # Prefer the versioned library if a specific NCCL version is specified
set(CMAKE_FIND_LIBRARY_SUFFIXES ".a.${NCCL_VERSION}" ${CMAKE_FIND_LIBRARY_SUFFIXES})
endif()
else()
SET(NCCL_LIBNAME "nccl")
if (NCCL_VERSION) # Prefer the versioned library if a specific NCCL version is specified
set(CMAKE_FIND_LIBRARY_SUFFIXES ".so.${NCCL_VERSION}" ${CMAKE_FIND_LIBRARY_SUFFIXES})
endif()
endif()
find_library(NCCL_LIBRARIES
NAMES ${NCCL_LIBNAME}
HINTS ${NCCL_LIB_DIR})
include(FindPackageHandleStandardArgs)
find_package_handle_standard_args(NCCL DEFAULT_MSG NCCL_INCLUDE_DIRS NCCL_LIBRARIES)
if(NCCL_FOUND) # obtaining NCCL version and some sanity checks
set (NCCL_HEADER_FILE "${NCCL_INCLUDE_DIRS}/nccl.h")
message (STATUS "Determining NCCL version from ${NCCL_HEADER_FILE}...")
set (OLD_CMAKE_REQUIRED_INCLUDES ${CMAKE_REQUIRED_INCLUDES})
list (APPEND CMAKE_REQUIRED_INCLUDES ${NCCL_INCLUDE_DIRS})
include(CheckCXXSymbolExists)
check_cxx_symbol_exists(NCCL_VERSION_CODE nccl.h NCCL_VERSION_DEFINED)
if (NCCL_VERSION_DEFINED)
set(file "${PROJECT_BINARY_DIR}/detect_nccl_version.cc")
file(WRITE ${file} "
#include <iostream>
#include <nccl.h>
int main()
{
std::cout << NCCL_MAJOR << '.' << NCCL_MINOR << '.' << NCCL_PATCH << std::endl;
int x;
ncclGetVersion(&x);
return x == NCCL_VERSION_CODE;
}
")
try_run(NCCL_VERSION_MATCHED compile_result ${PROJECT_BINARY_DIR} ${file}
RUN_OUTPUT_VARIABLE NCCL_VERSION_FROM_HEADER
CMAKE_FLAGS "-DINCLUDE_DIRECTORIES=${NCCL_INCLUDE_DIRS}"
LINK_LIBRARIES ${NCCL_LIBRARIES})
if (NOT NCCL_VERSION_MATCHED)
message(FATAL_ERROR "Found NCCL header version and library version do not match! \
(include: ${NCCL_INCLUDE_DIRS}, library: ${NCCL_LIBRARIES}) Please set NCCL_INCLUDE_DIR and NCCL_LIB_DIR manually.")
endif()
message(STATUS "NCCL version: ${NCCL_VERSION_FROM_HEADER}")
else()
message(STATUS "NCCL version < 2.3.5-5")
endif ()
set (CMAKE_REQUIRED_INCLUDES ${OLD_CMAKE_REQUIRED_INCLUDES})
message(STATUS "Found NCCL (include: ${NCCL_INCLUDE_DIRS}, library: ${NCCL_LIBRARIES})")
mark_as_advanced(NCCL_ROOT_DIR NCCL_INCLUDE_DIRS NCCL_LIBRARIES)
endif()