Fix dequantize_affine before iOS18 (#2589)

* Fix dequantize_affine before iOS18

* Update test_torch_quantization_ops.py

* fix ci

---------

Co-authored-by: yifan_shen3 <[email protected]>

Author: metascroy

coremltools/converters/mil/frontend/torch/quantization_ops.py

@@ -803,7 +803,6 @@ def dequantize_affine(context, node):
         int_data.astype(quantized_np_dtype),
         zero_point,
         scale,
-        axis=-1,
         name=node.name,
     )
     context.add(output, node.name)

coremltools/converters/mil/frontend/torch/test/test_torch_quantization_ops.py

@@ -272,6 +272,51 @@ def forward(self, x):
         prog = res[1]._mil_program
         assert get_op_types_in_program(prog) == ["constexpr_blockwise_shift_scale", "linear"]
 
+    @pytest.mark.skipif(not _HAS_TORCHAO, reason=MSG_TORCHAO_NOT_FOUND)
+    @pytest.mark.parametrize(
+        "compute_unit, has_zeros, minimum_deployment_target",
+        itertools.product(compute_units, [True, False], [ct.target.iOS16, ct.target.iOS17]),
+    )
+    def test_dequantize_affine_before_ios18(self, compute_unit, has_zeros, minimum_deployment_target):
+        quant_min = -128
+        quant_max = 127
+
+        n = 4
+        k = 128
+        input_dtype = torch.int8
+        int_data = torch.randint(low=quant_min, high=quant_max, size=(n, k)).to(input_dtype)
+        scale = torch.rand(n, 1)
+
+        zero_point = None
+        if has_zeros:
+            zero_point = torch.randint(low=quant_min, high=quant_max, size=(n, 1)).to(input_dtype)
+
+        class Model(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.register_buffer("int_data", int_data)
+                self.register_buffer("scale", scale)
+                self.register_buffer("zero_point", zero_point)
+
+            def forward(self, x):
+                w = torchao_quant.dequantize_affine(self.int_data, [1, k], self.scale, self.zero_point, input_dtype, quant_min, quant_max)
+                return torch.nn.functional.linear(x, w)
+
+
+        model = Model()
+        model = model.to(torch.device("cpu"))
+       
+        input_shape = [(3, k)]
+        res = self.run_compare_torch(
+            input_shape,
+            model,
+            minimum_deployment_target=minimum_deployment_target,
+            compute_unit=compute_unit,
+            rtol=0.1,
+            frontend=TorchFrontend.TORCHEXPORT,
+        )
+        prog = res[1]._mil_program
+        assert get_op_types_in_program(prog) == ["constexpr_affine_dequantize", "linear"]
 
 
 # TODO(rdar://108463675): refactor torch op tests later to parametrize quantized vs standard ops