diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
index a7bfb78a2f..f37a652f4d 100644
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -428,7 +428,7 @@ jobs:
           cache-name: cache-gpuocelot-build
         with:
           path: ${{ github.workspace }}/gpuocelot/ocelot
-          key: ubuntu22.04-gpuocelot-18401f4245b27ca4b3af433196583cc81ef84480-rebuild-2
+          key: ubuntu22.04-gpuocelot-18401f4245b27ca4b3af433196583cc81ef84480-rebuild-3
       - name: Clone/compile gpuocelot
         if: (matrix.backend == 'cuda' || matrix.backend == 'ptx' || matrix.backend == 'triton') && steps.cache-build.outputs.cache-hit != 'true'
         run: |
diff --git a/autogen_stubs.sh b/autogen_stubs.sh
index 922ac28e7e..82b71f8b23 100755
--- a/autogen_stubs.sh
+++ b/autogen_stubs.sh
@@ -15,10 +15,15 @@ fi
 
 BASE=tinygrad/runtime/autogen/
 
+fixup() {
+  sed -i '1s/^/# mypy: ignore-errors\n/' $1
+  sed -i 's/ *$//' $1
+  grep FIXME_STUB $1 || true
+}
+
 generate_opencl() {
   clang2py /usr/include/CL/cl.h -o $BASE/opencl.py -l /usr/lib/x86_64-linux-gnu/libOpenCL.so.1 -k cdefstum
-  sed -i '1s/^/# mypy: ignore-errors\n/' $BASE/opencl.py
-  grep FIXME_STUB $BASE/opencl.py || true
+  fixup $BASE/opencl.py
   # hot patches
   sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/opencl.py
   sed -i "s\ctypes.CDLL('/usr/lib/x86_64-linux-gnu/libOpenCL.so.1')\ctypes.CDLL(ctypes.util.find_library('OpenCL'))\g" $BASE/opencl.py
@@ -31,8 +36,7 @@ generate_hip() {
   --clang-args="-D__HIP_PLATFORM_AMD__ -I/opt/rocm/include -x c++" -o $BASE/hip.py -l /opt/rocm/lib/libamdhip64.so
   echo "hipDeviceProp_t = hipDeviceProp_tR0600" >> $BASE/hip.py
   echo "hipGetDeviceProperties = hipGetDevicePropertiesR0600" >> $BASE/hip.py
-  sed -i '1s/^/# mypy: ignore-errors\n/' $BASE/hip.py
-  grep FIXME_STUB $BASE/hip.py || true
+  fixup $BASE/hip.py
   # we can trust HIP is always at /opt/rocm/lib
   #sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/hip.py
   #sed -i "s\ctypes.CDLL('/opt/rocm/lib/libhiprtc.so')\ctypes.CDLL(ctypes.util.find_library('hiprtc'))\g" $BASE/hip.py
@@ -41,8 +45,7 @@ generate_hip() {
 
   clang2py /opt/rocm/include/amd_comgr/amd_comgr.h \
   --clang-args="-D__HIP_PLATFORM_AMD__ -I/opt/rocm/include -x c++" -o $BASE/comgr.py -l /opt/rocm/lib/libamd_comgr.so
-  sed -i '1s/^/# mypy: ignore-errors\n/' $BASE/comgr.py
-  grep FIXME_STUB $BASE/comgr.py || true
+  fixup $BASE/comgr.py
   python3 -c "import tinygrad.runtime.autogen.comgr"
 }
 
@@ -51,8 +54,7 @@ generate_cuda() {
   sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/cuda.py
   sed -i "s\ctypes.CDLL('/usr/lib/x86_64-linux-gnu/libcuda.so')\ctypes.CDLL(ctypes.util.find_library('cuda'))\g" $BASE/cuda.py
   sed -i "s\ctypes.CDLL('/usr/lib/x86_64-linux-gnu/libnvrtc.so')\ctypes.CDLL(ctypes.util.find_library('nvrtc'))\g" $BASE/cuda.py
-  sed -i '1s/^/# mypy: ignore-errors\n/' $BASE/cuda.py
-  grep FIXME_STUB $BASE/cuda.py || true
+  fixup $BASE/cuda.py
   python3 -c "import tinygrad.runtime.autogen.cuda"
 }
 
diff --git a/tinygrad/runtime/autogen/hip.py b/tinygrad/runtime/autogen/hip.py
index a50237dad2..b2e1a7de8d 100644
--- a/tinygrad/runtime/autogen/hip.py
+++ b/tinygrad/runtime/autogen/hip.py
@@ -131,7 +131,7 @@ class FunctionFactoryStub:
 
 # libraries['FIXME_STUB'] explanation
 # As you did not list (-l libraryname.so) a library that exports this function
-# This is a non-working stub instead. 
+# This is a non-working stub instead.
 # You can either re-run clan2py with -l /path/to/library.so
 # Or manually fix this by comment the ctypes.CDLL loading
 _libraries = {}
diff --git a/tinygrad/runtime/autogen/opencl.py b/tinygrad/runtime/autogen/opencl.py
index cd74a0869e..e2b73a8660 100644
--- a/tinygrad/runtime/autogen/opencl.py
+++ b/tinygrad/runtime/autogen/opencl.py
@@ -614,13 +614,13 @@ CL_VERSION_PATCH_BITS = (12) # macro
 # CL_VERSION_MINOR_MASK = ((1<<(10)) # macro
 # CL_VERSION_PATCH_MASK = ((1<<(12)) # macro
 # def CL_VERSION_MAJOR(version):  # macro
-#    return ((version)>>((10)+(12)))  
+#    return ((version)>>((10)+(12)))
 # def CL_VERSION_MINOR(version):  # macro
-#    return (((version)>>(12))&((1<<(10)))  
+#    return (((version)>>(12))&((1<<(10)))
 # def CL_VERSION_PATCH(version):  # macro
-#    return ((version)&((1<<(12)))  
+#    return ((version)&((1<<(12)))
 # def CL_MAKE_VERSION(major, minor, patch):  # macro
-#    return ((((major)&((1<<(10)))<<((10)+(12)))|(((minor)&((1<<(10)))<<(12))|((patch)&((1<<(12))))  
+#    return ((((major)&((1<<(10)))<<((10)+(12)))|(((minor)&((1<<(10)))<<(12))|((patch)&((1<<(12))))
 class struct__cl_platform_id(Structure):
     pass