diff --git a/llama/build_cuda.ps1 b/llama/build_cuda.ps1
deleted file mode 100644
index 736c722d..00000000
--- a/llama/build_cuda.ps1
+++ /dev/null
@@ -1,24 +0,0 @@
-nvcc -t 12 `
-    --generate-code=arch=compute_50,code=[compute_50,sm_50] `
-    --generate-code=arch=compute_52,code=[compute_52,sm_52] `
-    --generate-code=arch=compute_61,code=[compute_61,sm_61] `
-    --generate-code=arch=compute_70,code=[compute_70,sm_70] `
-    --generate-code=arch=compute_75,code=[compute_75,sm_75] `
-    --generate-code=arch=compute_80,code=[compute_80,sm_80] `
-    -DGGML_CUDA_DMMV_X=32 `
-    -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 `
-    -DGGML_CUDA_MMV_Y=1 `
-    -DGGML_USE_CUDA=1 `
-    -DGGML_SHARED=1 `
-    -DGGML_BUILD=1 `
-    -DGGML_USE_LLAMAFILE `
-    -Wno-deprecated-gpu-targets `
-    --forward-unknown-to-host-compiler `
-    -use_fast_math `
-    -link `
-    -shared `
-    -I. `
-    -lcuda -lcublas -lcudart -lcublasLt `
-    -O3 `
-    -o ggml-cuda.dll `
-    ggml-cuda.cu ggml-cuda/*.cu ggml.c ggml-backend.c ggml-alloc.c ggml-quants.c sgemm.cpp
diff --git a/llama/build_cuda.sh b/llama/build_cuda.sh
index e65ac4b1..c2f0ec03 100755
--- a/llama/build_cuda.sh
+++ b/llama/build_cuda.sh
@@ -1,3 +1,13 @@
+#!/bin/bash
+
+os="$(uname -s)"
+
+if [[ "$os" == "Windows_NT" || "$os" == "MINGW64_NT"* ]]; then
+    output="ggml-cuda.dll"
+else
+    output="libggml-cuda.so"
+fi
+
 nvcc \
     -t 12 \
     --generate-code=arch=compute_50,code=[compute_50,sm_50] \
@@ -14,6 +24,7 @@ nvcc \
     -DGGML_BUILD=1 \
     -DGGML_USE_LLAMAFILE \
     -D_GNU_SOURCE \
+    -DCMAKE_POSITION_INDEPENDENT_CODE=on \
     -Wno-deprecated-gpu-targets \
     --forward-unknown-to-host-compiler \
     -use_fast_math \
@@ -23,5 +34,5 @@ nvcc \
     -I. \
     -lcuda -lcublas -lcudart -lcublasLt \
     -O3 \
-    -o libggml-cuda.so \
+    -o $output \
     ggml-cuda.cu ggml-cuda/*.cu ggml.c ggml-backend.c ggml-alloc.c ggml-quants.c sgemm.cpp
diff --git a/llama/build_hipblas.sh b/llama/build_hipblas.sh
index a9a855df..f09106e6 100644
--- a/llama/build_hipblas.sh
+++ b/llama/build_hipblas.sh
@@ -1,16 +1,43 @@
+#!/bin/bash
+
+archs=(
+    gfx900
+    gfx940
+    gfx941
+    gfx942
+    gfx1010
+    gfx1012
+    gfx1030
+    gfx1100
+    gfx1101
+    gfx1102
+)
+
+linux_archs=(
+    gfx906:xnack-
+    gfx908:xnack-
+    gfx90a:xnack+
+    gfx90a:xnack-
+)
+
+os="$(uname -s)"
+
+if [[ "$os" == "Windows_NT" || "$os" == "MINGW64_NT"* ]]; then
+    output="ggml-hipblas.dll"
+else
+    output="libggml-hipblas.so"
+    archs+=("${linux_archs[@]}")
+fi
+
+offload_arch_flags=""
+for arch in "${archs[@]}"; do
+    offload_arch_flags+=" --offload-arch=$arch"
+done
+
 hipcc \
     -parallel-jobs=12 \
     -O3 \
-    --offload-arch=gfx900 \
-    --offload-arch=gfx940 \
-    --offload-arch=gfx941 \
-    --offload-arch=gfx942 \
-    --offload-arch=gfx1010 \
-    --offload-arch=gfx1012 \
-    --offload-arch=gfx1030 \
-    --offload-arch=gfx1100 \
-    --offload-arch=gfx1101 \
-    --offload-arch=gfx1102 \
+    $offload_arch_flags \
     -DGGML_USE_CUDA \
     -DGGML_BUILD=1 \
     -DGGML_SHARED=1 \
@@ -23,6 +50,7 @@ hipcc \
     -DNDEBUG \
     -DK_QUANTS_PER_ITERATION=2 \
     -D_CRT_SECURE_NO_WARNINGS \
+    -DCMAKE_POSITION_INDEPENDENT_CODE=on \
     -Xclang --dependent-lib=msvcrt -Wl,/subsystem:console \
     -Wno-expansion-to-defined \
     -Wno-invalid-noreturn \
@@ -35,10 +63,6 @@ hipcc \
     -o ggml-hipblas.dll \
     ggml-cuda.cu ggml-cuda/*.cu ggml.c ggml-backend.c ggml-alloc.c ggml-quants.c sgemm.cpp
 
-    # --offload-arch='gfx906:xnack-' \
-    # --offload-arch='gfx908:xnack-' \
-    # --offload-arch='gfx90a:xnack+' \
-    # --offload-arch='gfx90a:xnack-' \
     # -D_DLL \
     # -D_MT \
     # -D_XOPEN_SOURCE=600 \