up

Author: metascroy

torchao/experimental/tests/test_int8_dynamic_activation_intx_weight.py

@@ -360,7 +360,7 @@ def test_export_QDQLayout(self):
         self.assertTrue(torch.allclose(eager_results, exported_results))
 
         expected_lines = [
-            "torch.ops.torchao.choose_qparams_affine.default(input_1, 'ASYMMETRIC', [1, 512], torch.int8, None, None, None, torch.float32, torch.int32)",
+            "torch.ops.torchao.choose_qparams_affine.default(input_1, 'ASYMMETRIC', [1, 512], torch.int8, None, None, None, torch.float32, torch.int8)",
             "torch.ops.torchao.quantize_affine.default(input_1, [1, 512], getitem, getitem_1, torch.int8)",
             "torch.ops.torchao.dequantize_affine.default(quantize_affine, [1, 512], getitem, getitem_1, torch.int8)",
             "torch.ops.torchao.dequantize_affine.default",

torchao/quantization/quant_api.py

@@ -569,7 +569,7 @@ def _int8_asymm_per_token_quant(x: torch.Tensor) -> torch.Tensor:
     mapping_type = MappingType.ASYMMETRIC
     target_dtype = torch.int8
     scale_dtype = torch.float32
-    zero_point_dtype = torch.int32
+    zero_point_dtype = torch.int8
     if TORCH_VERSION_AT_LEAST_2_6:
         return to_affine_quantized_intx(
             x,