Adjust rtol value for confirming quantization.

Fixed for better readability.
This constraint should be enough to test whether a f32 value dequantized from i8 was changed.

PiperOrigin-RevId: 631559455

tensorflow/compiler/mlir/quantization/stablehlo/python/integration_test/quantize_model_test.py

@@ -899,7 +899,7 @@ def data_gen() -> repr_dataset.RepresentativeDataset:
     # be exactly the same. Indirectly proves that the `FunctionNameMatcherSpec`
     # with regex '.*invalid_function_name.*' did not match the quantizable unit.
     self.assertAllClose(new_outputs, expected_outputs, rtol=0.04)
-    self.assertNotAllClose(new_outputs, expected_outputs, rtol=0.00001)
+    self.assertNotAllClose(new_outputs, expected_outputs, 1e-7)
 
     # Due to other meta data, the compression is not exactly 1/4.
     self.assertLess(