Upstream version bump, and adapted `disable-rocm-cuda.gen_linux.sh.patch` accordingly.

3 files changed, 157 insertions, 41 deletions

diff --git a/.SRCINFO b/.SRCINFO
index 2f08a4c4cc01..0c5aa9208447 100644
--- a/.SRCINFO
+++ b/.SRCINFO
@@ -1,6 +1,6 @@
 pkgbase = ollama-nogpu-git
 	pkgdesc = Create, run and share large language models (LLMs). Package(s) without dedicated GPU offloading (no CUDA, no ROCm, no SYCL).
-	pkgver = 0.1.32+1.r2377.20240416.9df6c85c
+	pkgver = 0.1.37+3.r2707.20240511.4ec7445a
 	pkgrel = 1
 	url = https://github.com/jmorganca/ollama
 	arch = armv7h
@@ -26,7 +26,7 @@ pkgbase = ollama-nogpu-git
 	source = tmpfiles.d
 	b2sums = SKIP
 	b2sums = SKIP
-	b2sums = 490289e7afe8720792a7890636737147041a1f60d91f70f1b39a758f75e00ebc21724a1beb4d6ec74b0ff868636c887953f50fffbb998edc56ab087fe9477bbb
+	b2sums = 02d92a87554f2842b100908395a2599a843d21b95bf2c1961c2d862c35dda906893af107460054bf02340c52e4b4e3baa80cf95d071968fb9e158e697edaf0c6
 	b2sums = a773bbf16cf5ccc2ee505ad77c3f9275346ddf412be283cfeaee7c2e4c41b8637a31aaff8766ed769524ebddc0c03cf924724452639b62208e578d98b9176124
 	b2sums = 3aabf135c4f18e1ad745ae8800db782b25b15305dfeaaa031b4501408ab7e7d01f66e8ebb5be59fc813cfbff6788d08d2e48dcf24ecc480a40ec9db8dbce9fec
 	b2sums = e8f2b19e2474f30a4f984b45787950012668bf0acb5ad1ebb25cd9776925ab4a6aa927f8131ed53e35b1c71b32c504c700fe5b5145ecd25c7a8284373bb951ed
@@ -37,8 +37,8 @@ pkgname = ollama-generic-git
 	depends = glibc
 	depends = openssl
 	optdepends = 
-	provides = ollama=0.1.32+1.r2377.20240416.9df6c85c
-	provides = ollama-git=0.1.32+1.r2377.20240416.9df6c85c
+	provides = ollama=0.1.37+3.r2707.20240511.4ec7445a
+	provides = ollama-git=0.1.37+3.r2707.20240511.4ec7445a
 	conflicts = ollama
 
 pkgname = ollama-openmpi-git
@@ -47,8 +47,8 @@ pkgname = ollama-openmpi-git
 	depends = glibc
 	depends = openssl
 	optdepends = 
-	provides = ollama=0.1.32+1.r2377.20240416.9df6c85c
-	provides = ollama-git=0.1.32+1.r2377.20240416.9df6c85c
+	provides = ollama=0.1.37+3.r2707.20240511.4ec7445a
+	provides = ollama-git=0.1.37+3.r2707.20240511.4ec7445a
 	conflicts = ollama
 
 pkgname = ollama-openblas-git
@@ -57,8 +57,8 @@ pkgname = ollama-openblas-git
 	depends = glibc
 	depends = openssl
 	optdepends = 
-	provides = ollama=0.1.32+1.r2377.20240416.9df6c85c
-	provides = ollama-git=0.1.32+1.r2377.20240416.9df6c85c
+	provides = ollama=0.1.37+3.r2707.20240511.4ec7445a
+	provides = ollama-git=0.1.37+3.r2707.20240511.4ec7445a
 	conflicts = ollama
 
 pkgname = ollama-vulkan-git
@@ -67,6 +67,6 @@ pkgname = ollama-vulkan-git
 	depends = glibc
 	depends = openssl
 	optdepends = 
-	provides = ollama=0.1.32+1.r2377.20240416.9df6c85c
-	provides = ollama-git=0.1.32+1.r2377.20240416.9df6c85c
+	provides = ollama=0.1.37+3.r2707.20240511.4ec7445a
+	provides = ollama-git=0.1.37+3.r2707.20240511.4ec7445a
 	conflicts = ollama
diff --git a/PKGBUILD b/PKGBUILD
index 31e7b8f94260..54aff6270136 100644
--- a/PKGBUILD
+++ b/PKGBUILD
@@ -35,7 +35,7 @@ if "${_build_vulkan}"; then
   pkgname+=("${_name}-vulkan-git")
 fi
 pkgdesc='Create, run and share large language models (LLMs). Package(s) without dedicated GPU offloading (no CUDA, no ROCm, no SYCL).'
-pkgver=0.1.32+1.r2377.20240416.9df6c85c
+pkgver=0.1.37+3.r2707.20240511.4ec7445a
 pkgrel=1
 arch=(
   'armv7h'
@@ -79,15 +79,15 @@ source=(
   "tmpfiles.d"
 )
 b2sums=(
-  'SKIP'
-  'SKIP'
-  '490289e7afe8720792a7890636737147041a1f60d91f70f1b39a758f75e00ebc21724a1beb4d6ec74b0ff868636c887953f50fffbb998edc56ab087fe9477bbb'
-  'a773bbf16cf5ccc2ee505ad77c3f9275346ddf412be283cfeaee7c2e4c41b8637a31aaff8766ed769524ebddc0c03cf924724452639b62208e578d98b9176124'
-  '3aabf135c4f18e1ad745ae8800db782b25b15305dfeaaa031b4501408ab7e7d01f66e8ebb5be59fc813cfbff6788d08d2e48dcf24ecc480a40ec9db8dbce9fec'
-  'e8f2b19e2474f30a4f984b45787950012668bf0acb5ad1ebb25cd9776925ab4a6aa927f8131ed53e35b1c71b32c504c700fe5b5145ecd25c7a8284373bb951ed'
+  'SKIP'  # ollama (git)
+  'SKIP'  # llama.cpp (git)
+  '02d92a87554f2842b100908395a2599a843d21b95bf2c1961c2d862c35dda906893af107460054bf02340c52e4b4e3baa80cf95d071968fb9e158e697edaf0c6'  # disable-rocm-cuda.gen_linux.sh.patch
+  'a773bbf16cf5ccc2ee505ad77c3f9275346ddf412be283cfeaee7c2e4c41b8637a31aaff8766ed769524ebddc0c03cf924724452639b62208e578d98b9176124'  # ollama.service
+  '3aabf135c4f18e1ad745ae8800db782b25b15305dfeaaa031b4501408ab7e7d01f66e8ebb5be59fc813cfbff6788d08d2e48dcf24ecc480a40ec9db8dbce9fec'  # sysusers.conf
+  'e8f2b19e2474f30a4f984b45787950012668bf0acb5ad1ebb25cd9776925ab4a6aa927f8131ed53e35b1c71b32c504c700fe5b5145ecd25c7a8284373bb951ed'  # tmpfiles.d
 )
 options+=('emptydirs')
-#options+=('!lto') # openmpi variant fails to link without LTO.
+#options+=('!lto') # openmpi variant fails to link _without_ LTO.
 
 _check_cpufeature() {
   ## Checks if the host CPU supports the feature passed as argument "$1".
@@ -245,9 +245,9 @@ build() {
 
   if "${_build_clblas}"; then
     cd "${srcdir}/ollama-clblas"
-    
+
     export OLLAMA_CUSTOM_CPU_DEFS="${_cmake_options_common} ${_cmake_options_clblas}"
-    
+
     printf '\n'
     printf '%s\n' "   > Compiling clblas variant ..."
     printf '\n'
diff --git a/disable-rocm-cuda.gen_linux.sh.patch b/disable-rocm-cuda.gen_linux.sh.patch
index 2c4a04015449..859e0d8465e9 100644
--- a/disable-rocm-cuda.gen_linux.sh.patch
+++ b/disable-rocm-cuda.gen_linux.sh.patch
@@ -1,16 +1,16 @@
-diff -rU1 ollama.orig/llm/generate/gen_linux.sh ollama/llm/generate/gen_linux.sh
---- ollama.orig/llm/generate/gen_linux.sh	2024-04-17 15:02:16.439575013 +0200
-+++ ollama/llm/generate/gen_linux.sh	2024-04-17 15:16:15.656217455 +0200
+diff -rU1 ollama.orig/llm/generate/gen_linux.sh ollama.nocuda.norocm/llm/generate/gen_linux.sh
+--- ollama.orig/llm/generate/gen_linux.sh	2024-05-12 11:46:20.026663875 +0200
++++ ollama.nocuda.norocm/llm/generate/gen_linux.sh	2024-05-12 11:58:20.653330156 +0200
 @@ -53,3 +53,3 @@
  fi
 -COMMON_CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off"
 +COMMON_CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=on"
  source $(dirname $0)/gen_common.sh
-@@ -67,3 +67,3 @@
-         CMAKE_TARGETS="--target llama --target ggml"
--        CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
-+        CMAKE_DEFS="-DBUILD_SHARED_LIBS=on -DLLAMA_NATIVE=on ${CMAKE_DEFS}"
-         BUILD_DIR="../build/linux/${ARCH}_static"
+@@ -66,3 +66,3 @@
+     CMAKE_TARGETS="--target llama --target ggml"
+-    CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
++    CMAKE_DEFS="-DBUILD_SHARED_LIBS=on -DLLAMA_NATIVE=on ${CMAKE_DEFS}"
+     BUILD_DIR="../build/linux/${ARCH}_static"
 @@ -95,3 +95,3 @@
  
 -        COMMON_CPU_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off"
@@ -21,7 +21,7 @@ diff -rU1 ollama.orig/llm/generate/gen_linux.sh ollama/llm/generate/gen_linux.sh
 -            CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
 +            CMAKE_DEFS="${COMMON_CPU_DEFS} ${CMAKE_DEFS}"
              BUILD_DIR="../build/linux/${ARCH}/cpu"
-@@ -143,64 +143,2 @@
+@@ -143,71 +143,76 @@
  
 -# If needed, look for the default CUDA toolkit location
 -if [ -z "${CUDA_LIB_DIR}" ] && [ -d /usr/local/cuda/lib64 ]; then
@@ -47,14 +47,22 @@ diff -rU1 ollama.orig/llm/generate/gen_linux.sh ollama/llm/generate/gen_linux.sh
 -    fi
 -    if [ "${ARCH}" == "arm64" ]; then
 -        echo "ARM CPU detected - disabling unsupported AVX instructions"
--        
+-
 -        # ARM-based CPUs such as M1 and Tegra do not support AVX extensions.
 -        #
--        # CUDA compute < 6.0 lacks proper FP16 support on ARM. 
--        # Disabling has minimal performance effect while maintaining compatibility. 
+-        # CUDA compute < 6.0 lacks proper FP16 support on ARM.
+-        # Disabling has minimal performance effect while maintaining compatibility.
 -        ARM64_DEFS="-DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_CUDA_F16=off"
 -    fi
--    CMAKE_DEFS="-DLLAMA_CUDA=on -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS}"
+-    # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
+-    if [ -n "${OLLAMA_CUSTOM_CUDA_DEFS}" ]; then
+-        echo "OLLAMA_CUSTOM_CUDA_DEFS=\"${OLLAMA_CUSTOM_CUDA_DEFS}\""
+-        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
+-        echo "Building custom CUDA GPU"
+-    else
+-        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
+-    fi
+-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS}"
 -    BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
 -    EXTRA_LIBS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
 -    build
@@ -77,18 +85,88 @@ diff -rU1 ollama.orig/llm/generate/gen_linux.sh ollama/llm/generate/gen_linux.sh
 -        fi
 -    done
 -    compress
--
+ 
 -fi
--
++### CUDA stuff deactivated by patch:
+ 
 -if [ -z "${ROCM_PATH}" ]; then
 -    # Try the default location in case it exists
 -    ROCM_PATH=/opt/rocm
 -fi
--
- if [ -z "${CLBlast_DIR}" ]; then
-@@ -211,28 +149,2 @@
- fi
--
++# # If needed, look for the default CUDA toolkit location
++# if [ -z "${CUDA_LIB_DIR}" ] && [ -d /usr/local/cuda/lib64 ]; then
++#     CUDA_LIB_DIR=/usr/local/cuda/lib64
++# fi
++# 
++# # If needed, look for CUDA on Arch Linux
++# if [ -z "${CUDA_LIB_DIR}" ] && [ -d /opt/cuda/targets/x86_64-linux/lib ]; then
++#     CUDA_LIB_DIR=/opt/cuda/targets/x86_64-linux/lib
++# fi
++# 
++# # Allow override in case libcudart is in the wrong place
++# if [ -z "${CUDART_LIB_DIR}" ]; then
++#     CUDART_LIB_DIR="${CUDA_LIB_DIR}"
++# fi
++# 
++# if [ -d "${CUDA_LIB_DIR}" ]; then
++#     echo "CUDA libraries detected - building dynamic CUDA library"
++#     init_vars
++#     CUDA_MAJOR=$(ls "${CUDA_LIB_DIR}"/libcudart.so.* | head -1 | cut -f3 -d. || true)
++#     if [ -n "${CUDA_MAJOR}" ]; then
++#         CUDA_VARIANT=_v${CUDA_MAJOR}
++#     fi
++#     if [ "${ARCH}" == "arm64" ]; then
++#         echo "ARM CPU detected - disabling unsupported AVX instructions"
++# 
++#         # ARM-based CPUs such as M1 and Tegra do not support AVX extensions.
++#         #
++#         # CUDA compute < 6.0 lacks proper FP16 support on ARM.
++#         # Disabling has minimal performance effect while maintaining compatibility.
++#         ARM64_DEFS="-DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_CUDA_F16=off"
++#     fi
++#     # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
++#     if [ -n "${OLLAMA_CUSTOM_CUDA_DEFS}" ]; then
++#         echo "OLLAMA_CUSTOM_CUDA_DEFS=\"${OLLAMA_CUSTOM_CUDA_DEFS}\""
++#         CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
++#         echo "Building custom CUDA GPU"
++#     else
++#         CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
++#     fi
++#     CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS}"
++#     BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
++#     EXTRA_LIBS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
++#     build
++# 
++#     # Carry the CUDA libs as payloads to help reduce dependency burden on users
++#     #
++#     # TODO - in the future we may shift to packaging these separately and conditionally
++#     #        downloading them in the install script.
++#     DEPS="$(ldd ${BUILD_DIR}/bin/ollama_llama_server )"
++#     for lib in libcudart.so libcublas.so libcublasLt.so ; do
++#         DEP=$(echo "${DEPS}" | grep ${lib} | cut -f1 -d' ' | xargs || true)
++#         if [ -n "${DEP}" -a -e "${CUDA_LIB_DIR}/${DEP}" ]; then
++#             cp "${CUDA_LIB_DIR}/${DEP}" "${BUILD_DIR}/bin/"
++#         elif [ -e "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" ]; then
++#             cp "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" "${BUILD_DIR}/bin/"
++#         elif [ -e "${CUDART_LIB_DIR}/${lib}" ]; then
++#             cp -d ${CUDART_LIB_DIR}/${lib}* "${BUILD_DIR}/bin/"
++#         else
++#             cp -d "${CUDA_LIB_DIR}/${lib}*" "${BUILD_DIR}/bin/"
++#         fi
++#     done
++#     compress
++# 
++# fi
++
++### ROCm stuff deactivated by patch:
++
++# if [ -z "${ROCM_PATH}" ]; then
++#     # Try the default location in case it exists
++#     ROCM_PATH=/opt/rocm
++# fi
+ 
+@@ -220,33 +225,35 @@
+ 
 -if [ -d "${ROCM_PATH}" ]; then
 -    echo "ROCm libraries detected - building dynamic ROCm library"
 -    if [ -f ${ROCM_PATH}/lib/librocblas.so.*.*.????? ]; then
@@ -96,10 +174,17 @@ diff -rU1 ollama.orig/llm/generate/gen_linux.sh ollama/llm/generate/gen_linux.sh
 -    fi
 -    init_vars
 -    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DLLAMA_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
+-    # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
+-    if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
+-        echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""
+-        CMAKE_DEFS="${CMAKE_DEFS} ${OLLAMA_CUSTOM_ROCM_DEFS}"
+-        echo "Building custom ROCM GPU"
+-    fi
 -    BUILD_DIR="../build/linux/${ARCH}/rocm${ROCM_VARIANT}"
 -    EXTRA_LIBS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -Wl,-rpath,\$ORIGIN/../../rocm/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
 -    build
--
++### ROCm stuff deactivated by patch:
+ 
 -    # Record the ROCM dependencies
 -    rm -f "${BUILD_DIR}/bin/deps.txt"
 -    touch "${BUILD_DIR}/bin/deps.txt"
@@ -114,4 +199,35 @@ diff -rU1 ollama.orig/llm/generate/gen_linux.sh ollama/llm/generate/gen_linux.sh
 -    fi
 -    compress
 -fi
++# if [ -d "${ROCM_PATH}" ]; then
++#     echo "ROCm libraries detected - building dynamic ROCm library"
++#     if [ -f ${ROCM_PATH}/lib/librocblas.so.*.*.????? ]; then
++#         ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)
++#     fi
++#     init_vars
++#     CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DLLAMA_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
++#     # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
++#     if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
++#         echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""
++#         CMAKE_DEFS="${CMAKE_DEFS} ${OLLAMA_CUSTOM_ROCM_DEFS}"
++#         echo "Building custom ROCM GPU"
++#     fi
++#     BUILD_DIR="../build/linux/${ARCH}/rocm${ROCM_VARIANT}"
++#     EXTRA_LIBS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -Wl,-rpath,\$ORIGIN/../../rocm/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
++#     build
++# 
++#     # Record the ROCM dependencies
++#     rm -f "${BUILD_DIR}/bin/deps.txt"
++#     touch "${BUILD_DIR}/bin/deps.txt"
++#     for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e rocm -e amdgpu -e libtinfo ); do
++#         echo "${dep}" >> "${BUILD_DIR}/bin/deps.txt"
++#     done
++#     # bomb out if for some reason we didn't get a few deps
++#     if [ $(cat "${BUILD_DIR}/bin/deps.txt" | wc -l ) -lt 8 ] ; then
++#         cat "${BUILD_DIR}/bin/deps.txt"
++#         echo "ERROR: deps file short"
++#         exit 1
++#     fi
++#     compress
++# fi