Merge pull request #109 from danpetry/dpetry/fixes

Various fixes

recipe/0001-use-conda-ar-not-system.patch

@@ -0,0 +1,16 @@
+This fixes an issue where the build fails on linux-aarch64, because a system ar
+can't be found. Presumably it's using the system ar on other platforms, too,
+which isn't ideal.
+Index: nccl/src/Makefile
+===================================================================
+--- nccl.orig/src/Makefile	2024-05-16 16:58:16.522514975 -0500
++++ nccl/src/Makefile	2024-05-16 17:47:45.959416166 -0500
+@@ -83,7 +83,7 @@
+ $(LIBDIR)/$(STATICLIBTARGET): $(LIBOBJ) $(DEVMANIFEST)
+ 	@printf "Archiving  %-35s > %s\n" $(STATICLIBTARGET) $@
+ 	mkdir -p $(LIBDIR)
+-	ar cr $@ $(LIBOBJ) $$(cat $(DEVMANIFEST))
++	$(AR) cr $@ $(LIBOBJ) $$(cat $(DEVMANIFEST))
+
+ $(PKGDIR)/nccl.pc : nccl.pc.in
+ 	mkdir -p $(PKGDIR)

recipe/build.sh

@@ -1,26 +1,23 @@
 #!/bin/bash
 
-EXTRA_ARGS="CUDARTLIB=\"cudart_static\""
-
 if [[ "${cuda_compiler_version}" =~ 12.* ]]; then
-  EXTRA_ARGS="${EXTRA_ARGS} CUDA_HOME=\"${PREFIX}\" NVCC=\"${BUILD_PREFIX}/bin/nvcc\""
+  export CUDA_HOME=${BUILD_PREFIX}
 elif [[ "${cuda_compiler_version}" != "None" ]]; then
-  EXTRA_ARGS="${EXTRA_ARGS} CUDA_HOME=\"${CUDA_PATH}\""
+  export CUDA_HOME=${CUDA_PATH}
 fi
 
 if [[ $CONDA_BUILD_CROSS_COMPILATION == "1" ]]; then
     if [[ $target_platform == linux-aarch64 ]]; then
-        EXTRA_ARGS="${EXTRA_ARGS} CUDA_LIB=\"${CUDA_HOME}/targets/sbsa-linux/lib/\""
+        export CUDA_LIB=${CUDA_HOME}/targets/sbsa-linux/lib/
     elif [[ $target_platform == linux-ppc64le ]]; then
-        EXTRA_ARGS="${EXTRA_ARGS} CUDA_LIB=\"${CUDA_HOME}/targets/ppc64le-linux/lib/\""
+        export CUDA_LIB=${CUDA_HOME}/targets/ppc64le-linux/lib/
     else
         echo "not supported"
         exit -1
     fi
 fi
 
-# `eval` is needed here for proper `${...}` expansion
-eval make -j${CPU_COUNT} src.lib ${EXTRA_ARGS}
+make -j${CPU_COUNT} src.lib
 
 make install PREFIX="${PREFIX}"
 

recipe/meta.yaml

@@ -9,11 +9,10 @@ source:
   url: https://github.com/NVIDIA/nccl/archive/v{{ version }}.tar.gz
   sha256: 6b946b70a9d2d01871842cbd15ec56488d358abe9a0f3767e372fddc3e241ba7
   patches:
-    # Upstreaming w/PR: https://github.com/NVIDIA/nccl/pull/854
-    - 0001-Allow-custom-NVCC-path.patch
+    - 0001-use-conda-ar-not-system.patch
 
 build:
-  number: 0
+  number: 1
   skip: true  # [(not linux) or cuda_compiler_version in (undefined, "None", "10.2")]
   ignore_run_exports_from:
     # Ignore `cudatoolkit` dependency in CUDA 11 builds