diff --git a/tinygrad/codegen/late/devectorizer.py b/tinygrad/codegen/late/devectorizer.py
index 54b09f1c6c..dd4c4556e8 100644
--- a/tinygrad/codegen/late/devectorizer.py
+++ b/tinygrad/codegen/late/devectorizer.py
@@ -358,10 +358,14 @@ pm_reduce = PatternMatcher([
 
 # add loads
 
+def add_load(idx:UOp):
+  if isinstance(idx.dtype, PtrDType): return None
+  assert isinstance(idx.src[0].dtype, PtrDType), f"param is not PtrDType {idx.src[0].dtype}"
+  return idx.replace(dtype=idx.src[0].dtype).load(dtype=idx.dtype.base)
+
 pm_add_loads = PatternMatcher([
   # add loads to non ptr index
-  (UPat(Ops.INDEX, name="idx"), lambda idx: None if isinstance(idx.dtype, PtrDType) else
-    idx.replace(dtype=idx.src[0].dtype).load(dtype=idx.dtype.base)),
+  (UPat(Ops.INDEX, name="idx"), add_load),
   # remove loads from stores
   (UPat(Ops.STORE, src=(UPat(Ops.LOAD), UPat(name="val")), name="s"), lambda s,val: s.replace(src=(s.src[0].src[0], val))),
 ])
diff --git a/tinygrad/uop/ops.py b/tinygrad/uop/ops.py
index 509d58433d..c0603fa0ae 100644
--- a/tinygrad/uop/ops.py
+++ b/tinygrad/uop/ops.py
@@ -270,9 +270,9 @@ class UOp(OpMixin, metaclass=UOpMetaClass):
       # REDUCE with empty axis is passthrough (lowered form)
       case Ops.REDUCE if len(self.arg[1]) == 0:
         # these can mismatch if there's a horizonal reduce
-        if self.dtype.count > 1:
+        if self.src[0].dtype.count > 1:
           assert len(self.src[0]._shape) == 1
-          return (self.dtype.count,)
+          return () if self.dtype.count == 1 else (self.dtype.count,)
         return self.src[0]._shape
 
       # TODO: disallow shape changing bitcast