Bump ONNXRuntime to 1.29.0

abseil.sh

@@ -1,6 +1,6 @@
 package: abseil
 version: "%(tag_basename)s"
-tag: "20240722.0"
+tag: "20250814.0"
 requires:
   - "GCC-Toolchain:(?!osx)"
 license: Apache v2

autotools.sh

@@ -17,7 +17,10 @@ build_requires:
 #!/bin/bash -e
 
 unset CXXFLAGS
-unset CFLAGS
+# Not unset: GCC 14 turns implicit declarations, implicit int and mismatched
+# pointers into errors, and these sources predate that. gettext 0.20.1 calls
+# free() without <stdlib.h>, for one.
+export CFLAGS="-Wno-implicit-function-declaration -Wno-implicit-int -Wno-int-conversion -Wno-incompatible-pointer-types"
 export EMACS=no
 
 case $ARCHITECTURE in
@@ -81,22 +84,15 @@ pushd libtool*
   hash -r
 popd
 
-# Do not judge me. I am simply trying to float.
-# Apparently slc6 needs a different order compared
-# to the rest.
-case $ARCHITECTURE in
-  slc6*|ubuntu14*)
-    # automake -- requires: m4, autoconf, gettext
-    pushd automake*
-      $USE_AUTORECONF && [ -e bootstrap ] && sh ./bootstrap
-      ./configure --prefix $INSTALLROOT
-      make MAKEINFO=true ${JOBS+-j $JOBS}
-      make MAKEINFO=true install
-      hash -r
-    popd
-  ;;
-  *) ;;
-esac
+# automake -- requires: m4, autoconf. Must come before gettext, whose
+# autoreconf needs our aclocal rather than the host's.
+pushd automake*
+  if $USE_AUTORECONF && [ -e bootstrap ]; then sh ./bootstrap; fi
+  ./configure --prefix $INSTALLROOT
+  make MAKEINFO=true ${JOBS+-j $JOBS}
+  make MAKEINFO=true install
+  hash -r
+popd
 
 
 # gettext -- requires: nothing special
@@ -121,22 +117,6 @@ pushd gettext*
   hash -r
 popd
 
-# Do not judge me. I am simply trying to float.
-case $ARCHITECTURE in
-  slc6*|ubuntu14*) ;;
-  *)
-    # automake -- requires: m4, autoconf, gettext
-    pushd automake*
-      $USE_AUTORECONF && [ -e bootstrap ] && sh ./bootstrap
-      ./configure --prefix $INSTALLROOT
-      make MAKEINFO=true ${JOBS+-j $JOBS}
-      make MAKEINFO=true install
-      hash -r
-    popd
-  ;;
-esac
-
-
 # pkgconfig -- requires: nothing special
 pushd pkg-config*
   OLD_LDFLAGS="$LDFLAGS"

cudnn_frontend.sh

@@ -0,0 +1,21 @@
+package: cudnn_frontend
+version: "%(tag_basename)s"
+tag: v1.24.0
+license: MIT
+source: https://github.com/NVIDIA/cudnn-frontend
+build_requires:
+  - alibuild-recipe-tools
+---
+#!/bin/bash -e
+
+# ONNXRuntime's CUDA provider pulls this in with FetchContent, which aliBuild
+# forbids from downloading (-DFETCHCONTENT_FULLY_DISCONNECTED=ON). It is handed
+# $CUDNN_FRONTEND_ROOT as FETCHCONTENT_SOURCE_DIR_CUDNN_FRONTEND and calls
+# add_subdirectory() on it, so install the source tree rather than headers.
+# Version from onnxruntime's cmake/deps.txt at the tag we build.
+rsync -a --chmod=ug=rwX --exclude '**/.git' --delete --delete-excluded \
+      "$SOURCEDIR"/ "$INSTALLROOT"/
+
+MODULEDIR="$INSTALLROOT/etc/modulefiles"
+mkdir -p "$MODULEDIR"
+alibuild-generate-module --cmake > "$MODULEDIR/$PKGNAME"

cutlass.sh

@@ -0,0 +1,20 @@
+package: cutlass
+version: "%(tag_basename)s"
+tag: v4.4.2
+license: BSD-3-Clause
+source: https://github.com/NVIDIA/cutlass
+build_requires:
+  - alibuild-recipe-tools
+---
+#!/bin/bash -e
+
+# Source tree, not an install: ONNXRuntime receives $CUTLASS_ROOT as
+# FETCHCONTENT_SOURCE_DIR_CUTLASS and add_subdirectory()s it. See
+# cudnn_frontend.sh, which exists for the same reason.
+# Version from onnxruntime's cmake/deps.txt at the tag we build.
+rsync -a --chmod=ug=rwX --exclude '**/.git' --delete --delete-excluded \
+      "$SOURCEDIR"/ "$INSTALLROOT"/
+
+MODULEDIR="$INSTALLROOT/etc/modulefiles"
+mkdir -p "$MODULEDIR"
+alibuild-generate-module --cmake > "$MODULEDIR/$PKGNAME"

gpu-system.sh

@@ -202,7 +202,22 @@
         add_feature - miopen
     fi
 
-    if [[ $ALIBUILD_O2_FORCE_GPU_MIGRAPHX == 1 ]] || [[ $GPU_FEATURES =~ (^|-)"miopen"(-|_|$) && ${ALIBUILD_O2_FORCE_GPU_MIGRAPHX} != 0 && -d /opt/rocm/lib/migraphx ]]; then
+    # MIGraphX has to be new enough for the ONNXRuntime we build, not merely
+    # present: its provider references migraphx_shape_fp4x2_type, which appears
+    # in MIGraphX 7.1 (bf16 and fp8e5m2fnuz arrived in 6.4). On anything older
+    # every other file compiles and the build then dies on three enum cases, so
+    # gate on the API itself -- this turns MIGraphX back on by itself once the
+    # host is upgraded, with no disable left behind to forget about.
+    #
+    # ROCm 6.x keeps these headers in a self-contained prefix under lib/, which
+    # is on no default include path; onnxruntime.sh finds the prefix and passes
+    # it. Look in both places.
+    MIGRAPHX_C_API=
+    for _hdr in /opt/rocm/lib/migraphx/include/migraphx/migraphx.h \
+                /opt/rocm/include/migraphx/migraphx.h; do
+      if [[ -f $_hdr ]] && grep -q fp4x2 "$_hdr"; then MIGRAPHX_C_API=$_hdr; break; fi
+    done
+    if [[ $ALIBUILD_O2_FORCE_GPU_MIGRAPHX == 1 ]] || [[ $GPU_FEATURES =~ (^|-)"miopen"(-|_|$) && ${ALIBUILD_O2_FORCE_GPU_MIGRAPHX} != 0 && -n $MIGRAPHX_C_API ]]; then
       add_feature - migraphx
     fi
 
@@ -214,8 +229,20 @@
       add_feature - tensorrt
     fi
 
-    if [[ $ALIBUILD_O2_FORCE_GPU == "1" ]] && ! [[ $GPU_FEATURES =~ (^|-)"miopen"(-|_|$) && ${GPU_FEATURES} =~ (^|-)"migraphx"(-|_|$) && ${GPU_FEATURES} =~ (^|-)"cudnn"(-|_|$) && ${GPU_FEATURES} =~ (^|-)"tensorrt"(-|_|$) ]]; then
-      GPU_FEATURES="error-ALIBUILD_O2_FORCE_GPU=1 set, but not all ML libraries detected"
+    # MIGraphX is gated on the C API above, so it is legitimately absent when the
+    # host ROCm is older than the ONNXRuntime needs -- demand it only when it
+    # could have been enabled, otherwise FORCE_GPU=1 makes that state unreachable
+    # and even ALIBUILD_O2_FORCE_GPU_MIGRAPHX=0 cannot opt out.
+    _ml_required="miopen cudnn tensorrt"
+    if [[ ${ALIBUILD_O2_FORCE_GPU_MIGRAPHX} == 1 ]] || [[ ${ALIBUILD_O2_FORCE_GPU_MIGRAPHX} != 0 && -n $MIGRAPHX_C_API ]]; then
+      _ml_required="$_ml_required migraphx"
+    fi
+    _ml_missing=
+    for _ml in $_ml_required; do
+      [[ $GPU_FEATURES =~ (^|-)"$_ml"(-|_|$) ]] || _ml_missing="$_ml_missing $_ml"
+    done
+    if [[ $ALIBUILD_O2_FORCE_GPU == "1" && -n $_ml_missing ]]; then
+      GPU_FEATURES="error-ALIBUILD_O2_FORCE_GPU=1 set, but not all ML libraries detected:$_ml_missing"
       break
     fi
 

grpc.sh

@@ -1,6 +1,6 @@
 package: grpc
 version: "%(tag_basename)s"
-tag: v1.71.0
+tag: v1.74.0
 requires:
   - protobuf
   - c-ares

mesos.sh

@@ -1,6 +1,6 @@
 package: mesos
 version: v1.11.0
-tag: 1.11.0-alice5
+tag: 1.11.0-alice6
 source: https://github.com/AliceO2Group/mesos.git
 requires:
   - zlib

onnx.sh

@@ -1,7 +1,7 @@
 package: onnx
-version: "v1.17.0-alice2"
-tag: v1.17.0-alice2
-source: https://github.com/alisw/onnx
+version: "v1.22.0"
+tag: v1.22.0
+source: https://github.com/onnx/onnx
 license: ApacheV2+BSD2C+CC0
 requires:
   - "GCC-Toolchain:(?!osx)"

onnxruntime.sh

@@ -1,6 +1,6 @@
 package: ONNXRuntime
 version: "%(tag_basename)s"
-tag: v1.22.0
+tag: v1.29.0
 license: MIT
 source: https://github.com/microsoft/onnxruntime
 requires:
@@ -13,6 +13,8 @@ requires:
   - Eigen3
   - onnx
   - gpu-system
+  - "cudnn_frontend:(?!osx)"
+  - "cutlass:(?!osx)"
 build_requires:
   - date
   - safe_int
@@ -73,13 +75,21 @@ fi
 
 # Optional GPU features
 ### MIGraphX
-if [[ "$ORT_ROCM_BUILD" -eq 1 ]] && [[ ${O2_GPU_MIGRAPHX_AVAILABLE:-0} == 1 ]] && [[ -z "$ORT_MIGRAPHX_BUILD" ]]; then
-  ORT_MIGRAPHX_BUILD="0" # Disable for now, not working
+# Not gated on ORT_ROCM_BUILD: upstream removed the ROCm execution provider
+# after v1.22, so MIGraphX is the only remaining AMD path and has to stand on
+# its own. It needs hip and migraphx from the ROCm installation.
+if [[ ${O2_GPU_MIGRAPHX_AVAILABLE:-0} == 1 ]] && [[ -z "$ORT_MIGRAPHX_BUILD" ]]; then
+  ORT_MIGRAPHX_BUILD="1"
 elif [[ -z "$ORT_MIGRAPHX_BUILD" ]]; then
   ORT_MIGRAPHX_BUILD="0"
 fi
 ### TensorRT
-if [[ "$ORT_CUDA_BUILD" -eq 1 ]] && [[ ${O2_GPU_TENSORRT_AVAILABLE:-0} == 1 ]] && [[ -z "$ORT_TENSORRT_BUILD" ]]; then
+# Also gated on onnx_tensorrt being available. Its provider FetchContents that
+# source, which we build with FETCHCONTENT_FULLY_DISCONNECTED=ON and nothing in
+# alidist provides -- so on a TensorRT-capable host the configure step fails
+# outright. Gating on the root, as with cudnn_frontend and cutlass, turns it
+# back on by itself once a recipe supplies one.
+if [[ "$ORT_CUDA_BUILD" -eq 1 ]] && [[ ${O2_GPU_TENSORRT_AVAILABLE:-0} == 1 ]] && [[ -n "$ONNX_TENSORRT_ROOT" ]] && [[ -z "$ORT_TENSORRT_BUILD" ]]; then
   ORT_TENSORRT_BUILD="1"
 elif [[ -z "$ORT_TENSORRT_BUILD" ]]; then
   ORT_TENSORRT_BUILD="0"
@@ -93,6 +103,22 @@ export ORT_MIGRAPHX_BUILD=$ORT_MIGRAPHX_BUILD
 export ORT_TENSORRT_BUILD=$ORT_TENSORRT_BUILD
 EOF
 
+# MIGraphX installs into a self-contained prefix under lib/ on ROCm 6.x
+# (/opt/rocm/lib/migraphx/include/migraphx/version.h), which is on no default
+# include path: onnxruntime's provider then fails on <migraphx/version.h> even
+# though find_package(migraphx) succeeded. Find the prefix and pass it, rather
+# than assuming the headers sit beside ROCm's own.
+if [[ "$ORT_MIGRAPHX_BUILD" == 1 ]]; then
+  for _p in "${O2_GPU_ROCM_HOME:-/opt/rocm}/lib/migraphx" "${O2_GPU_ROCM_HOME:-/opt/rocm}"; do
+    if [[ -f "$_p/include/migraphx/version.h" ]]; then
+      MIGRAPHX_HOME=$_p
+      CXXFLAGS="$CXXFLAGS -isystem $_p/include"
+      break
+    fi
+  done
+  echo "MIGRAPHX_HOME=${MIGRAPHX_HOME:-<not found>}"
+fi
+
 echo "O2_GPU_ROCM_HOME=$O2_GPU_ROCM_HOME"
 echo "O2_GPU_CUDA_HOME=$O2_GPU_CUDA_HOME"
 
@@ -149,7 +175,9 @@ cmake "cmake"
       ${PROTOBUF_ROOT:+-DONNX_CUSTOM_PROTOC_EXECUTABLE=$PROTOBUF_ROOT/bin/protoc}                           \
       ${RE2_ROOT:+-DRE2_INCLUDE_DIR=${RE2_ROOT}/include}                                                    \
       ${BOOST_ROOT:+-DBOOST_INCLUDE_DIR=${BOOST_ROOT}/include}                                              \
+      ${BOOST_ROOT:+-DFETCHCONTENT_SOURCE_DIR_MP11=${BOOST_ROOT}}                                            \
       -Donnxruntime_USE_MIGRAPHX=${ORT_MIGRAPHX_BUILD}                                                      \
+      ${MIGRAPHX_HOME:+-DAMD_MIGRAPHX_HOME=${MIGRAPHX_HOME}}                                                \
       -Donnxruntime_USE_ROCM=${ORT_ROCM_BUILD}                                                              \
       -Donnxruntime_ROCM_HOME=${O2_GPU_ROCM_HOME}                                                           \
       -Donnxruntime_CUDA_HOME=${O2_GPU_CUDA_HOME}                                                           \
@@ -164,12 +192,13 @@ cmake "cmake"
       -DMSVC=OFF                                                                                            \
       -Donnxruntime_USE_CUDA=${ORT_CUDA_BUILD}                                                              \
       -Donnxruntime_USE_CUDA_NHWC_OPS=${ORT_CUDA_BUILD}                                                     \
-      -DFETCHCONTENT_SOURCE_DIR_CUDNN_FRONTEND=${CUDNN_FRONTEND_ROOT}                                       \
-      -DFETCHCONTENT_SOURCE_DIR_CUTLASS=${CUTLASS_ROOT}                                                     \
+      ${CUDNN_FRONTEND_ROOT:+-DFETCHCONTENT_SOURCE_DIR_CUDNN_FRONTEND=${CUDNN_FRONTEND_ROOT}}             \
+      ${CUTLASS_ROOT:+-DFETCHCONTENT_SOURCE_DIR_CUTLASS=${CUTLASS_ROOT}}                                   \
+      ${ONNX_TENSORRT_ROOT:+-DFETCHCONTENT_SOURCE_DIR_ONNX_TENSORRT=${ONNX_TENSORRT_ROOT}}           \
       -Donnxruntime_FUZZ_ENABLED=OFF                                                                        \
       -Donnxruntime_USE_FLASH_ATTENTION=OFF                                                                 \
       -Donnxruntime_USE_LEAN_ATTENTION=OFF                                                                  \
-      -Donnxruntime_USE_MEMORY_EFFICIENT_ATTENTION=OFF                                                      \
+      -Donnxruntime_USE_MEMORY_EFFICIENT_ATTENTION=ON                                                       \
       -DCMAKE_CUDA_FLAGS="${CXXFLAGS} -Wno-error=deprecated-enum-float-conversion -Wno-error -Wno-error=missing-requires -w" \
       -DCMAKE_HIP_FLAGS="${CXXFLAGS} -Wno-error=deprecated-enum-float-conversion -Wno-error -Wno-error=missing-requires -w" \
       -DCMAKE_CXX_FLAGS="${CXXFLAGS} -Wno-unknown-warning -Wno-unknown-warning-option -Wno-pass-failed -Wno-error=unused-but-set-variable -Wno-pass-failed=transform-warning -Wno-error=deprecated -Wno-error=maybe-uninitialized -Wno-error=deprecated-enum-enum-conversion -Wno-error -Wno-error=missing-requires -w" \

protobuf.sh

@@ -1,6 +1,6 @@
 package: protobuf
-version: v29.3
-tag: v29.3
+version: v31.1
+tag: v31.1
 source: https://github.com/protocolbuffers/protobuf
 license: Google BSD Like
 build_requires: