Fix/quant test precision for Cuda (#2467)

* fix(tests): relax precision for test_int8_wo_quant_save_load on ROCm

* fix(tests): relax precision for int8 wo quant save/load test

Author: petrex

test/quantization/test_quant_api.py

@@ -307,7 +307,7 @@ def api(model):
         res = m2(*example_inputs)
 
         # TODO: figure out why ROCm has a larger error
-        atol, rtol = (1e-2, 1e-2) if torch.version.hip else (0, 0)
+        atol, rtol = (1e-2, 1e-2) if torch.version.hip else (1e-7, 1e-5)
         torch.testing.assert_close(ref, res.cpu(), atol=atol, rtol=rtol)
 
     @unittest.skipIf(