Skip to content

Updated CK/AITER Cmake Build #666

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Micky774 wants to merge 3 commits into
base: dev
Choose a base branch
from
+81 −32
Open

Updated CK/AITER Cmake Build #666

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
704a1c6
Added incremental build for CK FA backend
Micky774 Jul 8, 2026
1067559
Update transformer_engine/common/ck_fused_attn/CMakeLists.txt
Micky774 Jul 9, 2026
117231e
Updated qola to cached ensure aiter commit
Micky774 Jul 16, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion 3rdparty/QoLA
Open in desktop
Submodule QoLA updated 1 files
+87 −7 qola/build_tools/submodule.py
111 changes: 80 additions & 31 deletions transformer_engine/common/ck_fused_attn/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -128,6 +128,10 @@ endif()
set(__AITER_CACHE_DIR "")
set(__AITER_MHA_PATH "")
set(__QOLA_INCLUDE_DIR "")
# When a kernel source-build is deferred to build time, this holds the name of
# the custom target that produces te_libmha_{fwd,bwd}.so. Empty when we consume
# prebuilt/user-supplied libs (nothing to build).
set(__CK_KERNELS_TARGET "")
if(NOT "$ENV{NVTE_CK_JIT}" STREQUAL "0")
set(__USE_CK_JIT TRUE)
else()
Expand All @@ -146,9 +150,32 @@ elseif(DEFINED ENV{NVTE_AITER_PREBUILT_BASE_URL})
endif()

if(__AITER_CACHE_DIR STREQUAL "")
# If not available or not requested, build from source via QoLA
# If not available or not requested, build from source via QoLA.
#
# The kernel build is expensive and its inputs (AITER pin, GPU archs, codegen
# patches, JIT scripts, tunable defaults) change rarely. Instead of running it
# at configure time -- which re-runs on every reconfigure, i.e. every
# `pip install` -- we emit it as a build-time custom command whose outputs are
# te_libmha_{fwd,bwd}.so in the build tree. Ninja then rebuilds the kernels
# only when a declared input changes and skips them entirely otherwise.
list(JOIN CMAKE_HIP_ARCHITECTURES ";" GPU_ARCHS_STR)
message(STATUS "[AITER-BUILD] Building AITER kernels for ${GPU_ARCHS_STR} via QoLA.")
message(STATUS "[AITER-BUILD] CK kernels will be built from source for ${GPU_ARCHS_STR} via QoLA (build-time target).")

set(__CK_KERNELS_DIR "${CMAKE_CURRENT_BINARY_DIR}/ck_kernels")
set(__CK_KERNELS_LIB "${__CK_KERNELS_DIR}/lib")
set(__CK_KERNELS_INCLUDE "${__CK_KERNELS_DIR}/include")
set(__CK_FWD_SO "${__CK_KERNELS_LIB}/te_libmha_fwd.so")
set(__CK_BWD_SO "${__CK_KERNELS_LIB}/te_libmha_bwd.so")

# Cache-busting marker: any change to the archs/tunables/pin rewrites this file
# (file(GENERATE) only touches it when the content actually differs), which in
# turn re-triggers the custom command below via its DEPENDS.
set(__CK_KEY_FILE "${__CK_KERNELS_DIR}/ck_build_key.txt")
file(GENERATE OUTPUT "${__CK_KEY_FILE}"
CONTENT "jit=${__USE_CK_JIT}\narchs=${GPU_ARCHS_STR}\nbf16=${CK_FUSED_ATTN_FLOAT_TO_BFLOAT16_DEFAULT}\nextra=$ENV{CK_JIT_EXTRA_CACHE_KEY}\naiter=${AITER_SHA}\n")

set(__CK_BUILD_DEPENDS "${__QOLA_MANIFEST}" "${__CK_KEY_FILE}")

if(__USE_CK_JIT)
message(STATUS "[AITER-BUILD] CK_JIT is enabled; will build CK kernels via CK_JIT.")
set(__CK_JIT_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/ck_jit")
Expand All @@ -158,7 +185,15 @@ if(__AITER_CACHE_DIR STREQUAL "")
else()
set(__CK_JIT_SOURCE_DIR "${CMAKE_CURRENT_LIST_DIR}/../../../3rdparty/ck_jit")
endif()
execute_process(
# Rebuild when any JIT orchestrator script/patch/runtime changes.
file(GLOB __CK_JIT_SCRIPTS CONFIGURE_DEPENDS

@ipanfilo ipanfilo Jul 16, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Will it be easier to just GLOB all files? Build with CK_JIT does not update repository tree but only temp and install directories

@Micky774 Micky774 Jul 20, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Updated

"${__CK_JIT_SOURCE_DIR}/*.py"
"${__CK_JIT_SOURCE_DIR}/*.sh"
"${__CK_JIT_SOURCE_DIR}/*.cpp"
"${__CK_JIT_SOURCE_DIR}/*.patch")
add_custom_command(
OUTPUT "${__CK_FWD_SO}" "${__CK_BWD_SO}"
COMMAND ${CMAKE_COMMAND} -E make_directory "${__CK_KERNELS_LIB}"
COMMAND ${Python_EXECUTABLE} "${__CK_JIT_SOURCE_DIR}/ck_jit_build.py" full
--with-qola
--qola-dir ${__QOLA_DIR}
Expand All @@ -167,43 +202,47 @@ if(__AITER_CACHE_DIR STREQUAL "")
--gpu-archs "${GPU_ARCHS_STR}"
--aiter-dir ${__AITER_SOURCE_DIR}
--tmp-dir "${__CK_JIT_BUILD_DIR}"
--install-dir ${AITER_MHA_INSTALL_DIR}
--install-dir "${__CK_KERNELS_LIB}"

@ipanfilo ipanfilo Jul 16, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why intermediate location is needed in this scheme, only because we don't know the destitation at the moment of CK_FWD/BWD_SO declaration?

@Micky774 Micky774 Jul 20, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It's mainly to separate out the build/install tree so that we have an easier time of cleaning the build for incremental work.

--jit-name "te_ck_jit"
RESULT_VARIABLE CKJIT_BUILD_RESULT
DEPENDS ${__CK_BUILD_DEPENDS} ${__CK_JIT_SCRIPTS}
COMMENT "[AITER-BUILD] Building CK fused-attn kernels (CK-JIT) for ${GPU_ARCHS_STR}"
VERBATIM
USES_TERMINAL
)
if(NOT CKJIT_BUILD_RESULT EQUAL 0)
message(FATAL_ERROR "[AITER-BUILD] CK-JIT build failed.")
endif()
# CK_JIT emits the libs straight into the install dir; nothing to cache.
set(__AITER_MHA_PATH ${AITER_MHA_INSTALL_DIR})
# QoLA public headers are exported next to the JIT libs during the build.
set(__QOLA_INCLUDE_DIR "${__QOLA_BUILD_DIR}/include")
else()
set(__QOLA_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/qola")
execute_process(
# Cache the freshly built libs+headers into the aiter prebuilt cache so
# downstream consumers (and future clean builds) see a self-contained tree.
get_default_aiter_cache_dir(__AITER_CACHE_DIR_FOR_QOLA)
add_custom_command(
OUTPUT "${__CK_FWD_SO}" "${__CK_BWD_SO}"
COMMAND ${CMAKE_COMMAND} -E make_directory "${__CK_KERNELS_LIB}"
COMMAND ${CMAKE_COMMAND} -E env "PYTHONPATH=${__QOLA_DIR}:$ENV{PYTHONPATH}"
${Python_EXECUTABLE} -m qola.cli build
--manifest ${__QOLA_MANIFEST}
--aiter-root ${__AITER_SOURCE_DIR}
--output-dir ${__QOLA_BUILD_DIR}
--arch "${GPU_ARCHS_STR}"
--skip-checkout
RESULT_VARIABLE QOLA_BUILD_RESULT
# Stage libs+headers into the kernels output dir consumed by ck_fused_attn.
COMMAND ${CMAKE_COMMAND} -E copy_directory "${__QOLA_BUILD_DIR}/lib" "${__CK_KERNELS_LIB}"
COMMAND ${CMAKE_COMMAND} -E copy_directory "${__QOLA_BUILD_DIR}/include" "${__CK_KERNELS_INCLUDE}"
# Populate the prebuilt cache for reuse across clean builds.
COMMAND ${CMAKE_COMMAND} -E copy_directory "${__QOLA_BUILD_DIR}/lib" "${__AITER_CACHE_DIR_FOR_QOLA}/lib"
COMMAND ${CMAKE_COMMAND} -E copy_directory "${__QOLA_BUILD_DIR}/include" "${__AITER_CACHE_DIR_FOR_QOLA}/include"
DEPENDS ${__CK_BUILD_DEPENDS}
COMMENT "[AITER-BUILD] Building CK fused-attn kernels (QoLA) for ${GPU_ARCHS_STR}"
VERBATIM
USES_TERMINAL
)
if(NOT QOLA_BUILD_RESULT EQUAL 0)
message(FATAL_ERROR "[AITER-BUILD] QoLA build failed.")
endif()

# Copy the final .so libs and exported public headers into the aiter
# prebuilt cache so downstream consumers see a self-contained tree.
get_default_aiter_cache_dir(__AITER_CACHE_DIR)
set(__QOLA_CACHE_LIB "${__AITER_CACHE_DIR}/lib")
file(MAKE_DIRECTORY ${__QOLA_CACHE_LIB})
file(GLOB __QOLA_BUILT_LIBS "${__QOLA_BUILD_DIR}/lib/*.so")
file(COPY ${__QOLA_BUILT_LIBS} DESTINATION ${__QOLA_CACHE_LIB})
file(COPY "${__QOLA_BUILD_DIR}/include" DESTINATION "${__AITER_CACHE_DIR}")
set(__AITER_MHA_PATH "${__QOLA_CACHE_LIB}")
set(__QOLA_INCLUDE_DIR "${__AITER_CACHE_DIR}/include")
set(__QOLA_INCLUDE_DIR "${__CK_KERNELS_INCLUDE}")
endif()

add_custom_target(ck_fused_attn_kernels DEPENDS "${__CK_FWD_SO}" "${__CK_BWD_SO}")
set(__CK_KERNELS_TARGET ck_fused_attn_kernels)
set(__AITER_MHA_PATH "${__CK_KERNELS_LIB}")
else()
message(STATUS "[AITER-BUILD] Using pre-built AITER from ${__AITER_CACHE_DIR}")
set(__AITER_MHA_PATH "${__AITER_CACHE_DIR}/lib")
Expand All @@ -226,10 +265,12 @@ set(CK_FUSED_ATTN_COMPILE_OPTIONS)
list(APPEND CK_FUSED_ATTN_COMPILE_OPTIONS
-DCK_TILE_FLOAT_TO_BFLOAT16_DEFAULT=${CK_FUSED_ATTN_FLOAT_TO_BFLOAT16_DEFAULT})

# Public QoLA headers ship alongside the .so libs in ${__AITER_MHA_PATH}/../include
# (emitted by qola.cli build, or copied from the QoLA build dir above for the
# source-build path).
if(NOT EXISTS "${__QOLA_INCLUDE_DIR}/qola_config.h")
# Public QoLA headers ship alongside the .so libs in ${__QOLA_INCLUDE_DIR}.
# When building from source they are produced by the ck_fused_attn_kernels
# target (see add_dependencies below), so they need not exist at configure time;
# the order-only dependency guarantees they are present before ck_fused_attn is
# compiled. Only assert existence when consuming prebuilt/user-supplied libs.
if(NOT __CK_KERNELS_TARGET AND NOT EXISTS "${__QOLA_INCLUDE_DIR}/qola_config.h")
message(FATAL_ERROR "Could not find QoLA public headers at ${__QOLA_INCLUDE_DIR}.")
endif()

Expand All @@ -245,7 +286,15 @@ target_link_libraries(ck_fused_attn PUBLIC ${ck_fused_attn_LINKER_LIBS})
target_compile_options(ck_fused_attn PRIVATE ${CK_FUSED_ATTN_COMPILE_OPTIONS})
set_target_properties(ck_fused_attn PROPERTIES INSTALL_RPATH "$ORIGIN")

if (NOT "${__AITER_MHA_PATH}" STREQUAL "${AITER_MHA_INSTALL_DIR}")
if (__CK_KERNELS_TARGET)
# Source build: order ck_fused_attn after the kernel libs+headers are produced,
# then install the freshly built artifacts from the build tree. The CK-JIT path
# also stages a ck_jit/ runtime tree next to the libs (absent for the QoLA path,
# hence OPTIONAL).
add_dependencies(ck_fused_attn ${__CK_KERNELS_TARGET})
install(FILES ${__CK_FWD_SO} ${__CK_BWD_SO} DESTINATION ${AITER_MHA_INSTALL_DIR})
install(DIRECTORY "${__CK_KERNELS_LIB}/ck_jit" DESTINATION ${AITER_MHA_INSTALL_DIR} OPTIONAL)
elseif (NOT "${__AITER_MHA_PATH}" STREQUAL "${AITER_MHA_INSTALL_DIR}")
install(FILES ${__AITER_MHA_PATH}/te_libmha_fwd.so ${__AITER_MHA_PATH}/te_libmha_bwd.so DESTINATION ${AITER_MHA_INSTALL_DIR})
endif()
install(TARGETS ck_fused_attn DESTINATION ${AITER_MHA_INSTALL_DIR})