From 0c17973ce981bb64884eb6bf5c292acb0a611267 Mon Sep 17 00:00:00 2001
From: Max Rakitin <max.rakitin@exabyte.io>
Date: Thu, 8 Feb 2024 21:48:14 -0500
Subject: [PATCH 01/14] Pytorch v1.13.1

---
 .github/workflows/build.yml | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index f557770..616e4ac 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -24,10 +24,10 @@ jobs:
           # - repo: https://github.com/kuelumbus/rdkit-pypi/
           #   name: rdkit-pypi
           #   version: "rdkit-pypi"
-          - repo: https://github.com/ssciwr/ipywidgets-jsonschema
-            name: ipywidgets-jsonschema
-            package-name: ipywidgets_jsonschema
-            version: "v1.1.0"
+          - repo: https://github.com/pytorch/pytorch
+            name: pytorch
+            package-name: pytorch
+            version: "v1.13.1"
 
     steps:
       - name: Check out the repo

From b3a742d93febd40ca6dd9219172a3b79777e3496 Mon Sep 17 00:00:00 2001
From: Max Rakitin <max.rakitin@exabyte.io>
Date: Thu, 8 Feb 2024 22:02:01 -0500
Subject: [PATCH 02/14] Enable more verbose output

---
 .github/workflows/build.yml | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 616e4ac..ad28bcc 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -56,7 +56,8 @@ jobs:
           set -vxeuo pipefail
           git clone --depth 1 --branch ${{ matrix.version }} ${{ matrix.repo }}
           cd ${{ matrix.name }}
-          pyodide build
+          export DEBUG=1
+          pyodide build -vv
           ls -la
           tree .
 

From 442dbb839aadbf07d47a1b8a64c9ed48868523d5 Mon Sep 17 00:00:00 2001
From: Max Rakitin <max.rakitin@exabyte.io>
Date: Thu, 8 Feb 2024 22:06:02 -0500
Subject: [PATCH 03/14] More config flags per
 https://github.com/pytorch/pytorch/blob/v1.13.1/CMakeLists.txt

---
 .github/workflows/build.yml | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index ad28bcc..a179955 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -57,7 +57,9 @@ jobs:
           git clone --depth 1 --branch ${{ matrix.version }} ${{ matrix.repo }}
           cd ${{ matrix.name }}
           export DEBUG=1
-          pyodide build -vv
+          export USE_CUDA=0
+          export CMAKE_BUILD_TYPE=Debug
+          pyodide build
           ls -la
           tree .
 

From ad4d171f5a2b3c84173fa28d807afbec593cc588 Mon Sep 17 00:00:00 2001
From: Max Rakitin <max.rakitin@exabyte.io>
Date: Thu, 8 Feb 2024 22:28:17 -0500
Subject: [PATCH 04/14] More diagnostics

---
 .github/workflows/build.yml | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index a179955..93bb61d 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -59,6 +59,8 @@ jobs:
           export DEBUG=1
           export USE_CUDA=0
           export CMAKE_BUILD_TYPE=Debug
+          ls -al /home/runner/work/build-pyodide/build-pyodide/pytorch/build/bin/
+          ls -la /home/runner/work/build-pyodide/build-pyodide/pytorch/build/bin/protoc.js-3.13.0.0
           pyodide build
           ls -la
           tree .

From 9e8c696b5a5cd7c696db8f11f7e020c0f89ecdc7 Mon Sep 17 00:00:00 2001
From: Max Rakitin <max.rakitin@exabyte.io>
Date: Thu, 8 Feb 2024 22:39:53 -0500
Subject: [PATCH 05/14] Use conda-forge build script

---
 .github/workflows/build.yml | 180 +++++++++++++++++++++++++++++++++++-
 1 file changed, 175 insertions(+), 5 deletions(-)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 93bb61d..6a68f1a 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -56,11 +56,181 @@ jobs:
           set -vxeuo pipefail
           git clone --depth 1 --branch ${{ matrix.version }} ${{ matrix.repo }}
           cd ${{ matrix.name }}
-          export DEBUG=1
-          export USE_CUDA=0
-          export CMAKE_BUILD_TYPE=Debug
-          ls -al /home/runner/work/build-pyodide/build-pyodide/pytorch/build/bin/
-          ls -la /home/runner/work/build-pyodide/build-pyodide/pytorch/build/bin/protoc.js-3.13.0.0
+
+          # CF build script: https://github.com/conda-forge/pytorch-cpu-feedstock/blob/main/recipe/build.sh
+          # remove pyproject.toml to avoid installing deps from pip
+          rm -rf pyproject.toml
+
+          # uncomment to debug cmake build
+          # export CMAKE_VERBOSE_MAKEFILE=1
+
+          export USE_NUMA=0
+          export USE_ITT=0
+          export CFLAGS="$(echo $CFLAGS | sed 's/-fvisibility-inlines-hidden//g')"
+          export CXXFLAGS="$(echo $CXXFLAGS | sed 's/-fvisibility-inlines-hidden//g')"
+          export LDFLAGS="$(echo $LDFLAGS | sed 's/-Wl,--as-needed//g')"
+          export LDFLAGS="$(echo $LDFLAGS | sed 's/-Wl,-dead_strip_dylibs//g')"
+          export LDFLAGS_LD="$(echo $LDFLAGS_LD | sed 's/-dead_strip_dylibs//g')"
+          if [[ "$c_compiler" == "clang" ]]; then
+              export CXXFLAGS="$CXXFLAGS -Wno-deprecated-declarations -Wno-unknown-warning-option -Wno-error=unused-command-line-argument"
+              export CFLAGS="$CFLAGS -Wno-deprecated-declarations -Wno-unknown-warning-option -Wno-error=unused-command-line-argument"
+          else
+              export CXXFLAGS="$CXXFLAGS -Wno-deprecated-declarations -Wno-error=maybe-uninitialized"
+              export CFLAGS="$CFLAGS -Wno-deprecated-declarations -Wno-error=maybe-uninitialized"
+          fi
+
+          # This is not correctly found for linux-aarch64 since pytorch 2.0.0 for some reason
+          export _GLIBCXX_USE_CXX11_ABI=1
+
+          # KINETO seems to require CUPTI and will look quite hard for it.
+          # CUPTI seems to cause trouble when users install a version of
+          # cudatoolkit different than the one specified at compile time.
+          # https://github.com/conda-forge/pytorch-cpu-feedstock/issues/135
+          export USE_KINETO=OFF
+
+          if [[ "$target_platform" == "osx-64" ]]; then
+            export CXXFLAGS="$CXXFLAGS -DTARGET_OS_OSX=1"
+            export CFLAGS="$CFLAGS -DTARGET_OS_OSX=1"
+          fi
+
+          # Dynamic libraries need to be lazily loaded so that torch
+          # can be imported on system without a GPU
+          LDFLAGS="${LDFLAGS//-Wl,-z,now/-Wl,-z,lazy}"
+
+          export CMAKE_GENERATOR=Ninja
+          export CMAKE_LIBRARY_PATH=$PREFIX/lib:$PREFIX/include:$CMAKE_LIBRARY_PATH
+          export CMAKE_PREFIX_PATH=$PREFIX
+          export CMAKE_BUILD_TYPE=Release
+
+          for ARG in $CMAKE_ARGS; do
+            if [[ "$ARG" == "-DCMAKE_"* ]]; then
+              cmake_arg=$(echo $ARG | cut -d= -f1)
+              cmake_arg=$(echo $cmake_arg| cut -dD -f2-)
+              cmake_val=$(echo $ARG | cut -d= -f2-)
+              printf -v $cmake_arg "$cmake_val"
+              export ${cmake_arg}
+            fi
+          done
+          unset CMAKE_INSTALL_PREFIX
+          export TH_BINARY_BUILD=1
+          export PYTORCH_BUILD_VERSION=$PKG_VERSION
+          export PYTORCH_BUILD_NUMBER=$PKG_BUILDNUM
+
+          export INSTALL_TEST=0
+          export BUILD_TEST=0
+
+          export USE_SYSTEM_SLEEF=1
+          # use our protobuf
+          export BUILD_CUSTOM_PROTOBUF=OFF
+          rm -rf $PREFIX/bin/protoc
+
+          if [[ "${target_platform}" != "${build_platform}" ]]; then
+              # It helps cross compiled builds without emulation support to complete
+              # Use BUILD PREFIX protoc instead of the one that is from the host platform
+              sed -i.bak \
+                  "s,IMPORTED_LOCATION_RELEASE .*/bin/protoc,IMPORTED_LOCATION_RELEASE \"${BUILD_PREFIX}/bin/protoc," \
+                  ${PREFIX}/lib/cmake/protobuf/protobuf-targets-release.cmake
+          fi
+
+          # I don't know where this folder comes from, but it's interfering with the build in osx-64
+          rm -rf $PREFIX/git
+
+          if [[ "$CONDA_BUILD_CROSS_COMPILATION" == 1 ]]; then
+              export COMPILER_WORKS_EXITCODE=0
+              export COMPILER_WORKS_EXITCODE__TRYRUN_OUTPUT=""
+          fi
+
+          export MAX_JOBS=${CPU_COUNT}
+
+          if [[ "$blas_impl" == "generic" ]]; then
+              # Fake openblas
+              export BLAS=OpenBLAS
+              sed -i.bak "s#FIND_LIBRARY.*#set(OpenBLAS_LIB ${PREFIX}/lib/liblapack${SHLIB_EXT} ${PREFIX}/lib/libcblas${SHLIB_EXT} ${PREFIX}/lib/libblas${SHLIB_EXT})#g" cmake/Modules/FindOpenBLAS.cmake
+          else
+              export BLAS=MKL
+          fi
+
+          if [[ "$PKG_NAME" == "pytorch" ]]; then
+            PIP_ACTION=install
+            sed "s/3.12/$PY_VER/g" build/CMakeCache.txt.orig > build/CMakeCache.txt
+          else
+            # For the main script we just build a wheel for so that the C++/CUDA
+            # parts are built. Then they are reused in each python version.
+            PIP_ACTION=wheel
+          fi
+
+          # MacOS build is simple, and will not be for CUDA
+          if [[ "$OSTYPE" == "darwin"* ]]; then
+              # Produce macOS builds with torch.distributed support.
+              # This is enabled by default on Linux, but disabled by default on macOS,
+              # because it requires an non-bundled compile-time dependency (libuv
+              # through gloo). This dependency is made available through meta.yaml, so
+              # we can override the default and set USE_DISTRIBUTED=1.
+              export USE_DISTRIBUTED=1
+
+              if [[ "$target_platform" == "osx-arm64" ]]; then
+                  # MKLDNN did not support on Apple M1 at the time support Apple M1
+                  # was added. Revisit later
+                  export USE_MKLDNN=0
+              fi
+          elif [[ ${cuda_compiler_version} != "None" ]]; then
+              # Even though cudnn is used for CUDA builds, it's good to enable
+              # for MKLDNN for CUDA builds when CUDA builds are used on a machine
+              # with no NVIDIA GPUs. However compilation fails with mkldnn and cuda enabled.
+              export USE_MKLDNN=OFF
+              export USE_CUDA=1
+              if [[ ${cuda_compiler_version} == 9.0* ]]; then
+                  export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;7.0+PTX"
+                  export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
+              elif [[ ${cuda_compiler_version} == 9.2* ]]; then
+                  export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0+PTX"
+                  export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
+              elif [[ ${cuda_compiler_version} == 10.* ]]; then
+                  export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5+PTX"
+                  export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
+              elif [[ ${cuda_compiler_version} == 11.0* ]]; then
+                  export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0+PTX"
+                  export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
+              elif [[ ${cuda_compiler_version} == 11.1 ]]; then
+                  export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0;8.6+PTX"
+                  export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
+              elif [[ ${cuda_compiler_version} == 11.2 ]]; then
+                  export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0;8.6+PTX"
+                  export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
+              elif [[ ${cuda_compiler_version} == 11.8 ]]; then
+                  export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0;8.6;8.9+PTX"
+                  export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
+              elif [[ ${cuda_compiler_version} == 12.0 ]]; then
+                  export TORCH_CUDA_ARCH_LIST="5.0;6.0;6.1;7.0;7.5;8.0;8.6;8.9;9.0+PTX"
+                  # $CUDA_HOME not set in CUDA 12.0. Using $PREFIX
+                  export CUDA_TOOLKIT_ROOT_DIR="${PREFIX}"
+              else
+                  echo "unsupported cuda version. edit build_pytorch.sh"
+                  exit 1
+              fi
+              export TORCH_NVCC_FLAGS="-Xfatbin -compress-all"
+              export NCCL_ROOT_DIR=$PREFIX
+              export NCCL_INCLUDE_DIR=$PREFIX/include
+              export USE_SYSTEM_NCCL=1
+              export USE_STATIC_NCCL=0
+              export USE_STATIC_CUDNN=0
+              export MAGMA_HOME="${PREFIX}"
+          else
+              if [[ "$target_platform" != *-64 ]]; then
+                # Breakpad seems to not work on aarch64 or ppc64le
+                # https://github.com/pytorch/pytorch/issues/67083
+                export USE_BREAKPAD=0
+              fi
+              # MKLDNN is an Apache-2.0 licensed library for DNNs and is used
+              # for CPU builds. Not to be confused with MKL.
+              export USE_MKLDNN=1
+              export USE_CUDA=0
+              export CMAKE_TOOLCHAIN_FILE="${RECIPE_DIR}/cross-linux.cmake"
+          fi
+
+          echo '${CXX}'=${CXX}
+          echo '${PREFIX}'=${PREFIX}
+          # End of CF build script
           pyodide build
           ls -la
           tree .

From ee59cbe96548409cb91ec310b4834703df12d924 Mon Sep 17 00:00:00 2001
From: Max Rakitin <mrakitin@users.noreply.github.com>
Date: Thu, 8 Feb 2024 23:08:56 -0500
Subject: [PATCH 06/14] Remove unbound vars check

---
 .github/workflows/build.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 6a68f1a..d71787d 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -53,7 +53,7 @@ jobs:
 
       - name: Build with Pyodide
         run: |
-          set -vxeuo pipefail
+          set -vxeo pipefail
           git clone --depth 1 --branch ${{ matrix.version }} ${{ matrix.repo }}
           cd ${{ matrix.name }}
 

From 9708117d70093fc715bad47b4e31bf21ac6a21ee Mon Sep 17 00:00:00 2001
From: Max Rakitin <max.rakitin@exabyte.io>
Date: Fri, 9 Feb 2024 07:53:15 -0500
Subject: [PATCH 07/14] Do not use cuda

---
 .github/workflows/build.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index d71787d..1a54b8a 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -224,10 +224,10 @@ jobs:
               # MKLDNN is an Apache-2.0 licensed library for DNNs and is used
               # for CPU builds. Not to be confused with MKL.
               export USE_MKLDNN=1
-              export USE_CUDA=0
               export CMAKE_TOOLCHAIN_FILE="${RECIPE_DIR}/cross-linux.cmake"
           fi
 
+          export USE_CUDA=0
           echo '${CXX}'=${CXX}
           echo '${PREFIX}'=${PREFIX}
           # End of CF build script

From 451d23657dd1f8737b091729519ccfb3aca4295b Mon Sep 17 00:00:00 2001
From: Max Rakitin <max.rakitin@exabyte.io>
Date: Tue, 13 Feb 2024 18:27:32 -0500
Subject: [PATCH 08/14] Remove some env vars

---
 .github/workflows/build.yml | 74 ++++---------------------------------
 1 file changed, 8 insertions(+), 66 deletions(-)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 1a54b8a..d3fa4a0 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -159,73 +159,15 @@ jobs:
             PIP_ACTION=wheel
           fi
 
-          # MacOS build is simple, and will not be for CUDA
-          if [[ "$OSTYPE" == "darwin"* ]]; then
-              # Produce macOS builds with torch.distributed support.
-              # This is enabled by default on Linux, but disabled by default on macOS,
-              # because it requires an non-bundled compile-time dependency (libuv
-              # through gloo). This dependency is made available through meta.yaml, so
-              # we can override the default and set USE_DISTRIBUTED=1.
-              export USE_DISTRIBUTED=1
-
-              if [[ "$target_platform" == "osx-arm64" ]]; then
-                  # MKLDNN did not support on Apple M1 at the time support Apple M1
-                  # was added. Revisit later
-                  export USE_MKLDNN=0
-              fi
-          elif [[ ${cuda_compiler_version} != "None" ]]; then
-              # Even though cudnn is used for CUDA builds, it's good to enable
-              # for MKLDNN for CUDA builds when CUDA builds are used on a machine
-              # with no NVIDIA GPUs. However compilation fails with mkldnn and cuda enabled.
-              export USE_MKLDNN=OFF
-              export USE_CUDA=1
-              if [[ ${cuda_compiler_version} == 9.0* ]]; then
-                  export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;7.0+PTX"
-                  export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
-              elif [[ ${cuda_compiler_version} == 9.2* ]]; then
-                  export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0+PTX"
-                  export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
-              elif [[ ${cuda_compiler_version} == 10.* ]]; then
-                  export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5+PTX"
-                  export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
-              elif [[ ${cuda_compiler_version} == 11.0* ]]; then
-                  export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0+PTX"
-                  export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
-              elif [[ ${cuda_compiler_version} == 11.1 ]]; then
-                  export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0;8.6+PTX"
-                  export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
-              elif [[ ${cuda_compiler_version} == 11.2 ]]; then
-                  export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0;8.6+PTX"
-                  export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
-              elif [[ ${cuda_compiler_version} == 11.8 ]]; then
-                  export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0;8.6;8.9+PTX"
-                  export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
-              elif [[ ${cuda_compiler_version} == 12.0 ]]; then
-                  export TORCH_CUDA_ARCH_LIST="5.0;6.0;6.1;7.0;7.5;8.0;8.6;8.9;9.0+PTX"
-                  # $CUDA_HOME not set in CUDA 12.0. Using $PREFIX
-                  export CUDA_TOOLKIT_ROOT_DIR="${PREFIX}"
-              else
-                  echo "unsupported cuda version. edit build_pytorch.sh"
-                  exit 1
-              fi
-              export TORCH_NVCC_FLAGS="-Xfatbin -compress-all"
-              export NCCL_ROOT_DIR=$PREFIX
-              export NCCL_INCLUDE_DIR=$PREFIX/include
-              export USE_SYSTEM_NCCL=1
-              export USE_STATIC_NCCL=0
-              export USE_STATIC_CUDNN=0
-              export MAGMA_HOME="${PREFIX}"
-          else
-              if [[ "$target_platform" != *-64 ]]; then
-                # Breakpad seems to not work on aarch64 or ppc64le
-                # https://github.com/pytorch/pytorch/issues/67083
-                export USE_BREAKPAD=0
-              fi
-              # MKLDNN is an Apache-2.0 licensed library for DNNs and is used
-              # for CPU builds. Not to be confused with MKL.
-              export USE_MKLDNN=1
-              export CMAKE_TOOLCHAIN_FILE="${RECIPE_DIR}/cross-linux.cmake"
+          if [[ "$target_platform" != *-64 ]]; then
+            # Breakpad seems to not work on aarch64 or ppc64le
+            # https://github.com/pytorch/pytorch/issues/67083
+            export USE_BREAKPAD=0
           fi
+          # MKLDNN is an Apache-2.0 licensed library for DNNs and is used
+          # for CPU builds. Not to be confused with MKL.
+          export USE_MKLDNN=1
+          export CMAKE_TOOLCHAIN_FILE="${RECIPE_DIR}/cross-linux.cmake"
 
           export USE_CUDA=0
           echo '${CXX}'=${CXX}

From eca7c1fd81c09c887b5c6136d98a1e4541bd0c15 Mon Sep 17 00:00:00 2001
From: Max Rakitin <max.rakitin@exabyte.io>
Date: Tue, 13 Feb 2024 19:13:04 -0500
Subject: [PATCH 09/14] Add CF build script

---
 scripts/build-pytorch.sh | 198 +++++++++++++++++++++++++++++++++++++++
 1 file changed, 198 insertions(+)
 create mode 100644 scripts/build-pytorch.sh

diff --git a/scripts/build-pytorch.sh b/scripts/build-pytorch.sh
new file mode 100644
index 0000000..ad92fe9
--- /dev/null
+++ b/scripts/build-pytorch.sh
@@ -0,0 +1,198 @@
+#!/bin/bash
+
+set -ex
+
+# remove pyproject.toml to avoid installing deps from pip
+rm -rf pyproject.toml
+
+# uncomment to debug cmake build
+# export CMAKE_VERBOSE_MAKEFILE=1
+
+export USE_NUMA=0
+export USE_ITT=0
+export CFLAGS="$(echo $CFLAGS | sed 's/-fvisibility-inlines-hidden//g')"
+export CXXFLAGS="$(echo $CXXFLAGS | sed 's/-fvisibility-inlines-hidden//g')"
+export LDFLAGS="$(echo $LDFLAGS | sed 's/-Wl,--as-needed//g')"
+export LDFLAGS="$(echo $LDFLAGS | sed 's/-Wl,-dead_strip_dylibs//g')"
+export LDFLAGS_LD="$(echo $LDFLAGS_LD | sed 's/-dead_strip_dylibs//g')"
+if [[ "$c_compiler" == "clang" ]]; then
+    export CXXFLAGS="$CXXFLAGS -Wno-deprecated-declarations -Wno-unknown-warning-option -Wno-error=unused-command-line-argument"
+    export CFLAGS="$CFLAGS -Wno-deprecated-declarations -Wno-unknown-warning-option -Wno-error=unused-command-line-argument"
+else
+    export CXXFLAGS="$CXXFLAGS -Wno-deprecated-declarations -Wno-error=maybe-uninitialized"
+    export CFLAGS="$CFLAGS -Wno-deprecated-declarations -Wno-error=maybe-uninitialized"
+fi
+
+# This is not correctly found for linux-aarch64 since pytorch 2.0.0 for some reason
+export _GLIBCXX_USE_CXX11_ABI=1
+
+# KINETO seems to require CUPTI and will look quite hard for it.
+# CUPTI seems to cause trouble when users install a version of
+# cudatoolkit different than the one specified at compile time.
+# https://github.com/conda-forge/pytorch-cpu-feedstock/issues/135
+export USE_KINETO=OFF
+
+if [[ "$target_platform" == "osx-64" ]]; then
+  export CXXFLAGS="$CXXFLAGS -DTARGET_OS_OSX=1"
+  export CFLAGS="$CFLAGS -DTARGET_OS_OSX=1"
+fi
+
+# Dynamic libraries need to be lazily loaded so that torch
+# can be imported on system without a GPU
+LDFLAGS="${LDFLAGS//-Wl,-z,now/-Wl,-z,lazy}"
+
+export CMAKE_GENERATOR=Ninja
+export CMAKE_LIBRARY_PATH=$PREFIX/lib:$PREFIX/include:$CMAKE_LIBRARY_PATH
+export CMAKE_PREFIX_PATH=$PREFIX
+export CMAKE_BUILD_TYPE=Release
+
+for ARG in $CMAKE_ARGS; do
+  if [[ "$ARG" == "-DCMAKE_"* ]]; then
+    cmake_arg=$(echo $ARG | cut -d= -f1)
+    cmake_arg=$(echo $cmake_arg| cut -dD -f2-)
+    cmake_val=$(echo $ARG | cut -d= -f2-)
+    printf -v $cmake_arg "$cmake_val"
+    export ${cmake_arg}
+  fi
+done
+unset CMAKE_INSTALL_PREFIX
+export TH_BINARY_BUILD=1
+export PYTORCH_BUILD_VERSION=$PKG_VERSION
+export PYTORCH_BUILD_NUMBER=$PKG_BUILDNUM
+
+export INSTALL_TEST=0
+export BUILD_TEST=0
+
+export USE_SYSTEM_SLEEF=1
+# use our protobuf
+export BUILD_CUSTOM_PROTOBUF=OFF
+rm -rf $PREFIX/bin/protoc
+
+if [[ "${target_platform}" != "${build_platform}" ]]; then
+    # It helps cross compiled builds without emulation support to complete
+    # Use BUILD PREFIX protoc instead of the one that is from the host platform
+    sed -i.bak \
+        "s,IMPORTED_LOCATION_RELEASE .*/bin/protoc,IMPORTED_LOCATION_RELEASE \"${BUILD_PREFIX}/bin/protoc," \
+        ${PREFIX}/lib/cmake/protobuf/protobuf-targets-release.cmake
+fi
+
+# I don't know where this folder comes from, but it's interfering with the build in osx-64
+rm -rf $PREFIX/git
+
+if [[ "$CONDA_BUILD_CROSS_COMPILATION" == 1 ]]; then
+    export COMPILER_WORKS_EXITCODE=0
+    export COMPILER_WORKS_EXITCODE__TRYRUN_OUTPUT=""
+fi
+
+export MAX_JOBS=${CPU_COUNT}
+
+if [[ "$blas_impl" == "generic" ]]; then
+    # Fake openblas
+    export BLAS=OpenBLAS
+    sed -i.bak "s#FIND_LIBRARY.*#set(OpenBLAS_LIB ${PREFIX}/lib/liblapack${SHLIB_EXT} ${PREFIX}/lib/libcblas${SHLIB_EXT} ${PREFIX}/lib/libblas${SHLIB_EXT})#g" cmake/Modules/FindOpenBLAS.cmake
+else
+    export BLAS=MKL
+fi
+
+if [[ "$PKG_NAME" == "pytorch" ]]; then
+  PIP_ACTION=install
+  sed "s/3.12/$PY_VER/g" build/CMakeCache.txt.orig > build/CMakeCache.txt
+else
+  # For the main script we just build a wheel for so that the C++/CUDA
+  # parts are built. Then they are reused in each python version.
+  PIP_ACTION=wheel
+fi
+
+# MacOS build is simple, and will not be for CUDA
+if [[ "$OSTYPE" == "darwin"* ]]; then
+    # Produce macOS builds with torch.distributed support.
+    # This is enabled by default on Linux, but disabled by default on macOS,
+    # because it requires an non-bundled compile-time dependency (libuv
+    # through gloo). This dependency is made available through meta.yaml, so
+    # we can override the default and set USE_DISTRIBUTED=1.
+    export USE_DISTRIBUTED=1
+
+    if [[ "$target_platform" == "osx-arm64" ]]; then
+        # MKLDNN did not support on Apple M1 at the time support Apple M1
+        # was added. Revisit later
+        export USE_MKLDNN=0
+    fi
+elif [[ ${cuda_compiler_version} != "None" ]]; then
+    # Even though cudnn is used for CUDA builds, it's good to enable
+    # for MKLDNN for CUDA builds when CUDA builds are used on a machine
+    # with no NVIDIA GPUs. However compilation fails with mkldnn and cuda enabled.
+    export USE_MKLDNN=OFF
+    export USE_CUDA=1
+    if [[ ${cuda_compiler_version} == 9.0* ]]; then
+        export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;7.0+PTX"
+        export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
+    elif [[ ${cuda_compiler_version} == 9.2* ]]; then
+        export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0+PTX"
+        export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
+    elif [[ ${cuda_compiler_version} == 10.* ]]; then
+        export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5+PTX"
+        export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
+    elif [[ ${cuda_compiler_version} == 11.0* ]]; then
+        export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0+PTX"
+        export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
+    elif [[ ${cuda_compiler_version} == 11.1 ]]; then
+        export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0;8.6+PTX"
+        export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
+    elif [[ ${cuda_compiler_version} == 11.2 ]]; then
+        export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0;8.6+PTX"
+        export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
+    elif [[ ${cuda_compiler_version} == 11.8 ]]; then
+        export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0;8.6;8.9+PTX"
+        export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
+    elif [[ ${cuda_compiler_version} == 12.0 ]]; then
+        export TORCH_CUDA_ARCH_LIST="5.0;6.0;6.1;7.0;7.5;8.0;8.6;8.9;9.0+PTX"
+        # $CUDA_HOME not set in CUDA 12.0. Using $PREFIX
+        export CUDA_TOOLKIT_ROOT_DIR="${PREFIX}"
+    else
+        echo "unsupported cuda version. edit build_pytorch.sh"
+        exit 1
+    fi
+    export TORCH_NVCC_FLAGS="-Xfatbin -compress-all"
+    export NCCL_ROOT_DIR=$PREFIX
+    export NCCL_INCLUDE_DIR=$PREFIX/include
+    export USE_SYSTEM_NCCL=1
+    export USE_STATIC_NCCL=0
+    export USE_STATIC_CUDNN=0
+    export MAGMA_HOME="${PREFIX}"
+else
+    if [[ "$target_platform" != *-64 ]]; then
+      # Breakpad seems to not work on aarch64 or ppc64le
+      # https://github.com/pytorch/pytorch/issues/67083
+      export USE_BREAKPAD=0
+    fi
+    # MKLDNN is an Apache-2.0 licensed library for DNNs and is used
+    # for CPU builds. Not to be confused with MKL.
+    export USE_MKLDNN=1
+    export USE_CUDA=0
+    export CMAKE_TOOLCHAIN_FILE="${RECIPE_DIR}/cross-linux.cmake"
+fi
+
+echo '${CXX}'=${CXX}
+echo '${PREFIX}'=${PREFIX}
+$PREFIX/bin/python -m pip $PIP_ACTION . --no-deps -vvv --no-clean \
+    | sed "s,${CXX},\$\{CXX\},g" \
+    | sed "s,${PREFIX},\$\{PREFIX\},g"
+
+if [[ "$PKG_NAME" == "libtorch" ]]; then
+  mkdir -p $SRC_DIR/dist
+  pushd $SRC_DIR/dist
+  wheel unpack ../torch-*.whl
+  pushd torch-*
+  mv torch/bin/* ${PREFIX}/bin
+  mv torch/lib/* ${PREFIX}/lib
+  mv torch/share/* ${PREFIX}/share
+  for f in ATen caffe2 tensorpipe torch c10; do
+    mv torch/include/$f ${PREFIX}/include/$f
+  done
+  rm ${PREFIX}/lib/libtorch_python.*
+  popd
+  popd
+
+  # Keep the original backed up to sed later
+  cp build/CMakeCache.txt build/CMakeCache.txt.orig
+fi

From 1c1c49bd272025684c0b238d1b2170b17e0219a6 Mon Sep 17 00:00:00 2001
From: Max Rakitin <max.rakitin@exabyte.io>
Date: Tue, 13 Feb 2024 19:17:27 -0500
Subject: [PATCH 10/14] Source compiler vars from CF build script

---
 .github/workflows/build.yml | 119 ++----------------------------------
 scripts/build-pytorch.sh    |  44 ++++++-------
 2 files changed, 26 insertions(+), 137 deletions(-)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index d3fa4a0..7151ad9 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -54,125 +54,14 @@ jobs:
       - name: Build with Pyodide
         run: |
           set -vxeo pipefail
-          git clone --depth 1 --branch ${{ matrix.version }} ${{ matrix.repo }}
-          cd ${{ matrix.name }}
-
-          # CF build script: https://github.com/conda-forge/pytorch-cpu-feedstock/blob/main/recipe/build.sh
-          # remove pyproject.toml to avoid installing deps from pip
-          rm -rf pyproject.toml
-
-          # uncomment to debug cmake build
-          # export CMAKE_VERBOSE_MAKEFILE=1
-
-          export USE_NUMA=0
-          export USE_ITT=0
-          export CFLAGS="$(echo $CFLAGS | sed 's/-fvisibility-inlines-hidden//g')"
-          export CXXFLAGS="$(echo $CXXFLAGS | sed 's/-fvisibility-inlines-hidden//g')"
-          export LDFLAGS="$(echo $LDFLAGS | sed 's/-Wl,--as-needed//g')"
-          export LDFLAGS="$(echo $LDFLAGS | sed 's/-Wl,-dead_strip_dylibs//g')"
-          export LDFLAGS_LD="$(echo $LDFLAGS_LD | sed 's/-dead_strip_dylibs//g')"
-          if [[ "$c_compiler" == "clang" ]]; then
-              export CXXFLAGS="$CXXFLAGS -Wno-deprecated-declarations -Wno-unknown-warning-option -Wno-error=unused-command-line-argument"
-              export CFLAGS="$CFLAGS -Wno-deprecated-declarations -Wno-unknown-warning-option -Wno-error=unused-command-line-argument"
-          else
-              export CXXFLAGS="$CXXFLAGS -Wno-deprecated-declarations -Wno-error=maybe-uninitialized"
-              export CFLAGS="$CFLAGS -Wno-deprecated-declarations -Wno-error=maybe-uninitialized"
-          fi
-
-          # This is not correctly found for linux-aarch64 since pytorch 2.0.0 for some reason
-          export _GLIBCXX_USE_CXX11_ABI=1
-
-          # KINETO seems to require CUPTI and will look quite hard for it.
-          # CUPTI seems to cause trouble when users install a version of
-          # cudatoolkit different than the one specified at compile time.
-          # https://github.com/conda-forge/pytorch-cpu-feedstock/issues/135
-          export USE_KINETO=OFF
-
-          if [[ "$target_platform" == "osx-64" ]]; then
-            export CXXFLAGS="$CXXFLAGS -DTARGET_OS_OSX=1"
-            export CFLAGS="$CFLAGS -DTARGET_OS_OSX=1"
-          fi
-
-          # Dynamic libraries need to be lazily loaded so that torch
-          # can be imported on system without a GPU
-          LDFLAGS="${LDFLAGS//-Wl,-z,now/-Wl,-z,lazy}"
-
-          export CMAKE_GENERATOR=Ninja
-          export CMAKE_LIBRARY_PATH=$PREFIX/lib:$PREFIX/include:$CMAKE_LIBRARY_PATH
-          export CMAKE_PREFIX_PATH=$PREFIX
-          export CMAKE_BUILD_TYPE=Release
 
-          for ARG in $CMAKE_ARGS; do
-            if [[ "$ARG" == "-DCMAKE_"* ]]; then
-              cmake_arg=$(echo $ARG | cut -d= -f1)
-              cmake_arg=$(echo $cmake_arg| cut -dD -f2-)
-              cmake_val=$(echo $ARG | cut -d= -f2-)
-              printf -v $cmake_arg "$cmake_val"
-              export ${cmake_arg}
-            fi
-          done
-          unset CMAKE_INSTALL_PREFIX
-          export TH_BINARY_BUILD=1
-          export PYTORCH_BUILD_VERSION=$PKG_VERSION
-          export PYTORCH_BUILD_NUMBER=$PKG_BUILDNUM
+          # Source the variables from the CF build script: https://github.com/conda-forge/pytorch-cpu-feedstock/blob/main/recipe/build.sh
+          source scripts/build-pytorch.sh
 
-          export INSTALL_TEST=0
-          export BUILD_TEST=0
-
-          export USE_SYSTEM_SLEEF=1
-          # use our protobuf
-          export BUILD_CUSTOM_PROTOBUF=OFF
-          rm -rf $PREFIX/bin/protoc
-
-          if [[ "${target_platform}" != "${build_platform}" ]]; then
-              # It helps cross compiled builds without emulation support to complete
-              # Use BUILD PREFIX protoc instead of the one that is from the host platform
-              sed -i.bak \
-                  "s,IMPORTED_LOCATION_RELEASE .*/bin/protoc,IMPORTED_LOCATION_RELEASE \"${BUILD_PREFIX}/bin/protoc," \
-                  ${PREFIX}/lib/cmake/protobuf/protobuf-targets-release.cmake
-          fi
-
-          # I don't know where this folder comes from, but it's interfering with the build in osx-64
-          rm -rf $PREFIX/git
-
-          if [[ "$CONDA_BUILD_CROSS_COMPILATION" == 1 ]]; then
-              export COMPILER_WORKS_EXITCODE=0
-              export COMPILER_WORKS_EXITCODE__TRYRUN_OUTPUT=""
-          fi
-
-          export MAX_JOBS=${CPU_COUNT}
-
-          if [[ "$blas_impl" == "generic" ]]; then
-              # Fake openblas
-              export BLAS=OpenBLAS
-              sed -i.bak "s#FIND_LIBRARY.*#set(OpenBLAS_LIB ${PREFIX}/lib/liblapack${SHLIB_EXT} ${PREFIX}/lib/libcblas${SHLIB_EXT} ${PREFIX}/lib/libblas${SHLIB_EXT})#g" cmake/Modules/FindOpenBLAS.cmake
-          else
-              export BLAS=MKL
-          fi
-
-          if [[ "$PKG_NAME" == "pytorch" ]]; then
-            PIP_ACTION=install
-            sed "s/3.12/$PY_VER/g" build/CMakeCache.txt.orig > build/CMakeCache.txt
-          else
-            # For the main script we just build a wheel for so that the C++/CUDA
-            # parts are built. Then they are reused in each python version.
-            PIP_ACTION=wheel
-          fi
+          git clone --depth 1 --branch ${{ matrix.version }} ${{ matrix.repo }}
+          cd ${{ matrix.name }}
 
-          if [[ "$target_platform" != *-64 ]]; then
-            # Breakpad seems to not work on aarch64 or ppc64le
-            # https://github.com/pytorch/pytorch/issues/67083
-            export USE_BREAKPAD=0
-          fi
-          # MKLDNN is an Apache-2.0 licensed library for DNNs and is used
-          # for CPU builds. Not to be confused with MKL.
-          export USE_MKLDNN=1
-          export CMAKE_TOOLCHAIN_FILE="${RECIPE_DIR}/cross-linux.cmake"
 
-          export USE_CUDA=0
-          echo '${CXX}'=${CXX}
-          echo '${PREFIX}'=${PREFIX}
-          # End of CF build script
           pyodide build
           ls -la
           tree .
diff --git a/scripts/build-pytorch.sh b/scripts/build-pytorch.sh
index ad92fe9..72a9ccd 100644
--- a/scripts/build-pytorch.sh
+++ b/scripts/build-pytorch.sh
@@ -174,25 +174,25 @@ fi
 
 echo '${CXX}'=${CXX}
 echo '${PREFIX}'=${PREFIX}
-$PREFIX/bin/python -m pip $PIP_ACTION . --no-deps -vvv --no-clean \
-    | sed "s,${CXX},\$\{CXX\},g" \
-    | sed "s,${PREFIX},\$\{PREFIX\},g"
-
-if [[ "$PKG_NAME" == "libtorch" ]]; then
-  mkdir -p $SRC_DIR/dist
-  pushd $SRC_DIR/dist
-  wheel unpack ../torch-*.whl
-  pushd torch-*
-  mv torch/bin/* ${PREFIX}/bin
-  mv torch/lib/* ${PREFIX}/lib
-  mv torch/share/* ${PREFIX}/share
-  for f in ATen caffe2 tensorpipe torch c10; do
-    mv torch/include/$f ${PREFIX}/include/$f
-  done
-  rm ${PREFIX}/lib/libtorch_python.*
-  popd
-  popd
-
-  # Keep the original backed up to sed later
-  cp build/CMakeCache.txt build/CMakeCache.txt.orig
-fi
+# $PREFIX/bin/python -m pip $PIP_ACTION . --no-deps -vvv --no-clean \
+#     | sed "s,${CXX},\$\{CXX\},g" \
+#     | sed "s,${PREFIX},\$\{PREFIX\},g"
+
+# if [[ "$PKG_NAME" == "libtorch" ]]; then
+#   mkdir -p $SRC_DIR/dist
+#   pushd $SRC_DIR/dist
+#   wheel unpack ../torch-*.whl
+#   pushd torch-*
+#   mv torch/bin/* ${PREFIX}/bin
+#   mv torch/lib/* ${PREFIX}/lib
+#   mv torch/share/* ${PREFIX}/share
+#   for f in ATen caffe2 tensorpipe torch c10; do
+#     mv torch/include/$f ${PREFIX}/include/$f
+#   done
+#   rm ${PREFIX}/lib/libtorch_python.*
+#   popd
+#   popd
+
+#   # Keep the original backed up to sed later
+#   cp build/CMakeCache.txt build/CMakeCache.txt.orig
+# fi

From ee73af79c1f9a93e9ca08fea893c766477b22d90 Mon Sep 17 00:00:00 2001
From: Max Rakitin <max.rakitin@exabyte.io>
Date: Tue, 13 Feb 2024 19:18:54 -0500
Subject: [PATCH 11/14] Rearrange the order of sourcing and use v2.1.2

---
 .github/workflows/build.yml | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 7151ad9..49ada81 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -27,7 +27,7 @@ jobs:
           - repo: https://github.com/pytorch/pytorch
             name: pytorch
             package-name: pytorch
-            version: "v1.13.1"
+            version: "v2.1.2"
 
     steps:
       - name: Check out the repo
@@ -55,12 +55,11 @@ jobs:
         run: |
           set -vxeo pipefail
 
-          # Source the variables from the CF build script: https://github.com/conda-forge/pytorch-cpu-feedstock/blob/main/recipe/build.sh
-          source scripts/build-pytorch.sh
-
           git clone --depth 1 --branch ${{ matrix.version }} ${{ matrix.repo }}
           cd ${{ matrix.name }}
 
+          # Source the variables from the CF build script: https://github.com/conda-forge/pytorch-cpu-feedstock/blob/main/recipe/build.sh
+          source ../scripts/build-pytorch.sh
 
           pyodide build
           ls -la

From c0802c8fdad810cf6699f54f88a9da6d6d256fdc Mon Sep 17 00:00:00 2001
From: Max Rakitin <max.rakitin@exabyte.io>
Date: Tue, 13 Feb 2024 19:21:07 -0500
Subject: [PATCH 12/14] Print discovered cuda version

---
 scripts/build-pytorch.sh | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/scripts/build-pytorch.sh b/scripts/build-pytorch.sh
index 72a9ccd..92d8c12 100644
--- a/scripts/build-pytorch.sh
+++ b/scripts/build-pytorch.sh
@@ -123,6 +123,8 @@ elif [[ ${cuda_compiler_version} != "None" ]]; then
     # with no NVIDIA GPUs. However compilation fails with mkldnn and cuda enabled.
     export USE_MKLDNN=OFF
     export USE_CUDA=1
+
+    echo "cuda_compiler_version = ${cuda_compiler_version}"
     if [[ ${cuda_compiler_version} == 9.0* ]]; then
         export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;7.0+PTX"
         export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME

From f7f6ab45e86771f85f760231435043218d833672 Mon Sep 17 00:00:00 2001
From: Max Rakitin <max.rakitin@exabyte.io>
Date: Tue, 13 Feb 2024 19:23:23 -0500
Subject: [PATCH 13/14] Do not deal with cuda versions

---
 scripts/build-pytorch.sh | 44 ----------------------------------------
 1 file changed, 44 deletions(-)

diff --git a/scripts/build-pytorch.sh b/scripts/build-pytorch.sh
index 92d8c12..3590887 100644
--- a/scripts/build-pytorch.sh
+++ b/scripts/build-pytorch.sh
@@ -117,50 +117,6 @@ if [[ "$OSTYPE" == "darwin"* ]]; then
         # was added. Revisit later
         export USE_MKLDNN=0
     fi
-elif [[ ${cuda_compiler_version} != "None" ]]; then
-    # Even though cudnn is used for CUDA builds, it's good to enable
-    # for MKLDNN for CUDA builds when CUDA builds are used on a machine
-    # with no NVIDIA GPUs. However compilation fails with mkldnn and cuda enabled.
-    export USE_MKLDNN=OFF
-    export USE_CUDA=1
-
-    echo "cuda_compiler_version = ${cuda_compiler_version}"
-    if [[ ${cuda_compiler_version} == 9.0* ]]; then
-        export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;7.0+PTX"
-        export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
-    elif [[ ${cuda_compiler_version} == 9.2* ]]; then
-        export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0+PTX"
-        export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
-    elif [[ ${cuda_compiler_version} == 10.* ]]; then
-        export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5+PTX"
-        export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
-    elif [[ ${cuda_compiler_version} == 11.0* ]]; then
-        export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0+PTX"
-        export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
-    elif [[ ${cuda_compiler_version} == 11.1 ]]; then
-        export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0;8.6+PTX"
-        export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
-    elif [[ ${cuda_compiler_version} == 11.2 ]]; then
-        export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0;8.6+PTX"
-        export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
-    elif [[ ${cuda_compiler_version} == 11.8 ]]; then
-        export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0;8.6;8.9+PTX"
-        export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
-    elif [[ ${cuda_compiler_version} == 12.0 ]]; then
-        export TORCH_CUDA_ARCH_LIST="5.0;6.0;6.1;7.0;7.5;8.0;8.6;8.9;9.0+PTX"
-        # $CUDA_HOME not set in CUDA 12.0. Using $PREFIX
-        export CUDA_TOOLKIT_ROOT_DIR="${PREFIX}"
-    else
-        echo "unsupported cuda version. edit build_pytorch.sh"
-        exit 1
-    fi
-    export TORCH_NVCC_FLAGS="-Xfatbin -compress-all"
-    export NCCL_ROOT_DIR=$PREFIX
-    export NCCL_INCLUDE_DIR=$PREFIX/include
-    export USE_SYSTEM_NCCL=1
-    export USE_STATIC_NCCL=0
-    export USE_STATIC_CUDNN=0
-    export MAGMA_HOME="${PREFIX}"
 else
     if [[ "$target_platform" != *-64 ]]; then
       # Breakpad seems to not work on aarch64 or ppc64le

From f54c4148f9c397cf75ff175e5154bc7e0654753d Mon Sep 17 00:00:00 2001
From: Max Rakitin <max.rakitin@exabyte.io>
Date: Tue, 19 Mar 2024 01:54:09 -0400
Subject: [PATCH 14/14] Install `pyyaml` build dep

---
 .github/workflows/build.yml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 49ada81..54c2f7c 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -27,7 +27,7 @@ jobs:
           - repo: https://github.com/pytorch/pytorch
             name: pytorch
             package-name: pytorch
-            version: "v2.1.2"
+            version: "v1.13.1"
 
     steps:
       - name: Check out the repo
@@ -61,7 +61,7 @@ jobs:
           # Source the variables from the CF build script: https://github.com/conda-forge/pytorch-cpu-feedstock/blob/main/recipe/build.sh
           source ../scripts/build-pytorch.sh
 
-          pyodide build
+          pyodide build --build-dependencies pyyaml
           ls -la
           tree .