reproduced difference between means in python and c++ entirely in python still dont know why

.gitignore

@@ -1,3 +1,3 @@
-build
+build/
 .vscode
-data
+data/

src/main.cu

@@ -3,14 +3,16 @@
 #include <cuda_runtime.h>
 #include <device_launch_parameters.h>
 #include <iostream>
+#include <iomanip> 
 #include <cmath>
 
+
 int main() {
     std::string path = "data/MERRA2_400.inst6_3d_ana_Np.20120101.nc4";
     std::string variable = "T";
     auto x = readData(path, variable);
 
-    // Print some values from the file to see that it worked
+    // Calculate the mean of the data to see if it works.
     float sum = 0;
     int n = 0;
     int skipped = 0;

test_read.py

@@ -10,7 +10,7 @@ print(ncfile.variables.keys())
 
 U = ncfile.variables['T'][:] 
 
-# Check the shape of the variable (it should be 3D)
+# Check the shape of the variable 
 print("Shape of U:", U.shape)
 
 # Compute the mean of the variable across all axes (for all elements in U)
@@ -28,5 +28,19 @@ print("Number of masked values in U:", masked_count)
 nan_count = np.isnan(U).sum()
 print("Number of NaN values in U:", nan_count)
 
+print("Calculating mean manually (takes a bit cause python is slowww)")
+
+count = 0
+valsum = 0
+for val in U.flat:
+    if not np.ma.is_masked(val):
+        # print(val)
+        valsum += val
+        count += 1
+
+print(f"{valsum=} {valsum/count=} {count=}")
+
+print(f"The problem is this: why does {valsum/count=} not equal {U_mean=}")
+
 # Close the NetCDF file
 ncfile.close()