diff --git a/docs/development.md b/docs/development.md
index ac45a3e0..59651b1f 100644
--- a/docs/development.md
+++ b/docs/development.md
@@ -50,7 +50,8 @@ development and runtime packages.
 Typically the build scripts will auto-detect CUDA, however, if your Linux distro
 or installation approach uses unusual paths, you can specify the location by
 specifying an environment variable `CUDA_LIB_DIR` to the location of the shared
-libraries, and `CUDACXX` to the location of the nvcc compiler.
+libraries, and `CUDACXX` to the location of the nvcc compiler.  You can customize
+set set of target CUDA architectues by setting `CMAKE_CUDA_ARCHITECTURES` (e.g. "50;60;70")
 
 Then generate dependencies:
 
diff --git a/gpu/gpu.go b/gpu/gpu.go
index e234ec39..550467a3 100644
--- a/gpu/gpu.go
+++ b/gpu/gpu.go
@@ -30,8 +30,8 @@ type handles struct {
 var gpuMutex sync.Mutex
 var gpuHandles *handles = nil
 
-// With our current CUDA compile flags, 5.2 and older will not work properly
-const CudaComputeMajorMin = 6
+// With our current CUDA compile flags, older than 5.0 will not work properly
+var CudaComputeMin = [2]C.int{5, 0}
 
 // Possible locations for the nvidia-ml library
 var CudaLinuxGlobs = []string{
@@ -142,7 +142,7 @@ func GetGPUInfo() GpuInfo {
 			if cc.err != nil {
 				slog.Info(fmt.Sprintf("error looking up CUDA GPU compute capability: %s", C.GoString(cc.err)))
 				C.free(unsafe.Pointer(cc.err))
-			} else if cc.major >= CudaComputeMajorMin {
+			} else if cc.major > CudaComputeMin[0] || (cc.major == CudaComputeMin[0] && cc.minor >= CudaComputeMin[1]) {
 				slog.Info(fmt.Sprintf("CUDA Compute Capability detected: %d.%d", cc.major, cc.minor))
 				resp.Library = "cuda"
 			} else {
diff --git a/llm/generate/gen_common.sh b/llm/generate/gen_common.sh
index 0b7b2ae4..43d3dce5 100644
--- a/llm/generate/gen_common.sh
+++ b/llm/generate/gen_common.sh
@@ -39,6 +39,9 @@ init_vars() {
     *)
         ;;
     esac
+    if [ -z "${CMAKE_CUDA_ARCHITECTURES}" ] ; then 
+        CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
+    fi
 }
 
 git_module_setup() {
diff --git a/llm/generate/gen_linux.sh b/llm/generate/gen_linux.sh
index fee4c85f..82c8c75b 100755
--- a/llm/generate/gen_linux.sh
+++ b/llm/generate/gen_linux.sh
@@ -135,7 +135,7 @@ if [ -d "${CUDA_LIB_DIR}" ]; then
     if [ -n "${CUDA_MAJOR}" ]; then
         CUDA_VARIANT=_v${CUDA_MAJOR}
     fi
-    CMAKE_DEFS="-DLLAMA_CUBLAS=on ${COMMON_CMAKE_DEFS} ${CMAKE_DEFS}"
+    CMAKE_DEFS="-DLLAMA_CUBLAS=on -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${COMMON_CMAKE_DEFS} ${CMAKE_DEFS}"
     BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/cuda${CUDA_VARIANT}"
     EXTRA_LIBS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
     build
diff --git a/llm/generate/gen_windows.ps1 b/llm/generate/gen_windows.ps1
index e4271997..f7a241cc 100644
--- a/llm/generate/gen_windows.ps1
+++ b/llm/generate/gen_windows.ps1
@@ -25,6 +25,11 @@ function init_vars {
     }
     $script:GZIP=(get-command -ea 'silentlycontinue' gzip).path
     $script:DUMPBIN=(get-command -ea 'silentlycontinue' dumpbin).path
+    if ($null -eq $env:CMAKE_CUDA_ARCHITECTURES) {
+        $script:CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
+    } else {
+        $script:CMAKE_CUDA_ARCHITECTURES=$env:CMAKE_CUDA_ARCHITECTURES
+    }
 }
 
 function git_module_setup {
@@ -151,7 +156,7 @@ if ($null -ne $script:CUDA_LIB_DIR) {
     }
     init_vars
     $script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cuda$script:CUDA_VARIANT"
-    $script:cmakeDefs += @("-DLLAMA_CUBLAS=ON", "-DLLAMA_AVX=on")
+    $script:cmakeDefs += @("-DLLAMA_CUBLAS=ON", "-DLLAMA_AVX=on", "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}")
     build
     install
     cp "${script:CUDA_LIB_DIR}/cudart64_*.dll" "${script:buildDir}/lib"