From 4091e877bf09d15296dccdb12ac720bc7206d387 Mon Sep 17 00:00:00 2001
From: Daniel Petry <dpetry@anaconda.com>
Date: Thu, 16 May 2024 13:26:10 -0500
Subject: [PATCH 1/4] Don't build on s390x, it's unsupported by CUDA

---
 recipe/meta.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/recipe/meta.yaml b/recipe/meta.yaml
index 963a697..fdf809b 100644
--- a/recipe/meta.yaml
+++ b/recipe/meta.yaml
@@ -14,7 +14,7 @@ source:
 
 build:
   number: 0
-  skip: true  # [(not linux) or cuda_compiler_version in (undefined, "None", "10.2")]
+  skip: true  # [(not linux) or s390x or cuda_compiler_version in (undefined, "None", "10.2")]
   ignore_run_exports_from:
     # Ignore `cudatoolkit` dependency in CUDA 11 builds
     - {{ compiler("cuda") }}  # [(cuda_compiler_version or "").startswith("11")]

From 9ba7185f636fca301789fcdf9fe47ecd4b9c7b7c Mon Sep 17 00:00:00 2001
From: Daniel Petry <dpetry@anaconda.com>
Date: Thu, 16 May 2024 13:26:21 -0500
Subject: [PATCH 2/4] Build with CUDA v12.4

---
 recipe/conda_build_config.yaml | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/recipe/conda_build_config.yaml b/recipe/conda_build_config.yaml
index 1ea99f3..f3beb68 100644
--- a/recipe/conda_build_config.yaml
+++ b/recipe/conda_build_config.yaml
@@ -6,10 +6,11 @@
 #   - 7.3               # [linux]
 # cxx_compiler_version: # [linux]
 #   - 7.3               # [linux]
-cuda_compiler: nvcc
+cuda_compiler: cuda-nvcc
 cuda_compiler_version:
-  - 9.2
+  #- 9.2
   #- 10.0
   # - 10.1
   #- 10.2
-  #- 11.0
\ No newline at end of file
+  #- 11.0
+  - 12.4

From 3269d16fd2649e64d14b347a46562959fe34fe84 Mon Sep 17 00:00:00 2001
From: Daniel Petry <dpetry@anaconda.com>
Date: Thu, 16 May 2024 17:55:57 -0500
Subject: [PATCH 3/4] Make conda-build find the CUDA folder correctly

---
 recipe/build.sh  | 5 +++--
 recipe/meta.yaml | 4 ++--
 2 files changed, 5 insertions(+), 4 deletions(-)

diff --git a/recipe/build.sh b/recipe/build.sh
index ce8df31..0770377 100755
--- a/recipe/build.sh
+++ b/recipe/build.sh
@@ -4,6 +4,7 @@ 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}\""
 fi
@@ -19,8 +20,8 @@ if [[ $CONDA_BUILD_CROSS_COMPILATION == "1" ]]; then
     fi
 fi
 
-# `eval` is needed here for proper `${...}` expansion
-eval make -j${CPU_COUNT} src.lib ${EXTRA_ARGS}
+# Handing CUDA_HOME to make via an environment variable works; using eval and make args doesn't.
+make -j${CPU_COUNT} src.lib
 
 make install PREFIX="${PREFIX}"
 
diff --git a/recipe/meta.yaml b/recipe/meta.yaml
index fdf809b..06c2cd0 100644
--- a/recipe/meta.yaml
+++ b/recipe/meta.yaml
@@ -9,8 +9,8 @@ source:
   url: https://github.com/NVIDIA/nccl/archive/v{{ version }}.tar.gz
   sha256: 1923596984d85e310b5b6c52b2c72a1b93da57218f2bc5a5c7ac3d59297a3303
   patches:
-    # Upstreaming w/PR: https://github.com/NVIDIA/nccl/pull/854
-    - 0001-Allow-custom-NVCC-path.patch
+    # This isn't needed; setting CUDA_HOME sets NVCC correctly
+    #- 0001-Allow-custom-NVCC-path.patch
 
 build:
   number: 0

From c4c66c71ca2654bb42e0f8739a1e917b7de60fca Mon Sep 17 00:00:00 2001
From: Daniel Petry <dpetry@anaconda.com>
Date: Thu, 16 May 2024 17:56:48 -0500
Subject: [PATCH 4/4] Use conda ar rather than system one

---
 recipe/0002-use-conda-ar-not-system.patch | 16 ++++++++++++++++
 recipe/meta.yaml                          |  2 ++
 2 files changed, 18 insertions(+)
 create mode 100644 recipe/0002-use-conda-ar-not-system.patch

diff --git a/recipe/0002-use-conda-ar-not-system.patch b/recipe/0002-use-conda-ar-not-system.patch
new file mode 100644
index 0000000..0399e15
--- /dev/null
+++ b/recipe/0002-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)
diff --git a/recipe/meta.yaml b/recipe/meta.yaml
index 06c2cd0..8bc8bf4 100644
--- a/recipe/meta.yaml
+++ b/recipe/meta.yaml
@@ -11,6 +11,7 @@ source:
   patches:
     # This isn't needed; setting CUDA_HOME sets NVCC correctly
     #- 0001-Allow-custom-NVCC-path.patch
+    - 0002-use-conda-ar-not-system.patch
 
 build:
   number: 0
@@ -28,6 +29,7 @@ requirements:
     - {{ compiler("cxx") }}
     - {{ compiler("cuda") }}
     - make
+    - patch
   host:
     - cuda-version ={{ cuda_compiler_version }}          # [(cuda_compiler_version or "").startswith("11")]
   run: