[converter] Fix negative axis in quantize/dequantize lowering (#24)

* [converter] Fix negative axis in quantize/dequantize lowering

The quantize/dequantize lowering normalized a negative axis as
axis + rank - 1, off by one from the eager op, which resolves it as
axis + rank. A per-channel axis=-1 was applied one dimension early, so
the converted model used the wrong channel; when the channel and a
neighbor dim share a size this is silent, with no shape error.

Normalize a negative axis as axis + rank, matching the eager op. Add
per-channel negative-axis numerical tests for both ops. They use a
(2, 4, 4) shape with equal middle and channel dims so a wrong axis
surfaces as a numerical mismatch rather than a reshape error.

* Parametrize negative-axis tests over equal and distinct dims

---------

Co-authored-by: Henry Tao <55294647+jakesabathia2@users.noreply.github.com>

Author: eyupcanakman

coreai_torch/_custom_to_core.py

@@ -295,11 +295,11 @@ def _replace_quantize_or_dequantize(
         quant_elem_type = input_type.element_type  # dequantize: quant→float
         float_elem_type = result_elem_type
 
-    # Extract axis; normalize negative axis the same way the C++ lowering does.
+    # Extract axis; normalize a negative axis the same way the eager op does.
     axis_val = _get_optional_int_arg(node, axis_idx, default=0)
     input_rank = len(input_type.shape)
     if axis_val < 0:
-        axis_val = axis_val + input_rank - 1
+        axis_val = axis_val + input_rank
 
     axis = coreai.constant(np.array(axis_val, dtype=np.int32), loc=loc)
 

tests/ops/test_custom_ops.py

@@ -411,6 +411,30 @@ def forward(self, x: Tensor) -> Tensor:
             prepare_program=inject_subbyte_tensors,
         )
 
+    # (2, 4, 4): equal dims keep a wrong axis silent (a value mismatch).
+    # (2, 3, 4): distinct dims show -1 is the last dim (a wrong axis is a reshape error).
+    @pytest.mark.parametrize("x", [torch.randn(2, 4, 4), torch.randn(2, 3, 4)])
+    async def test_per_channel_negative_axis_numerical(self, x: Tensor) -> None:
+        """quantize with a per-channel scale on a negative axis matches eager."""
+
+        class Model(nn.Module):
+            def __init__(self) -> None:
+                super().__init__()
+                self.register_buffer(
+                    "scale", torch.tensor([0.1, 0.2, 0.3, 0.4], dtype=torch.float32)
+                )
+                self.register_buffer("zero_point", torch.zeros(4, dtype=torch.int8))
+
+            def forward(self, x: Tensor) -> Tensor:
+                return torch.ops.coreai.quantize(
+                    x, self.scale, torch.int8, zero_point=self.zero_point, axis=-1
+                )
+
+        model = Model()
+        await validate_numerical_output(
+            model=model, x=x, prepare_program=inject_subbyte_tensors
+        )
+
 
 # ---------------------------------------------------------------------------
 # dequantize → coreai.dequantize
@@ -540,6 +564,36 @@ def forward(self, x: Tensor) -> Tensor:
             prepare_program=inject_subbyte_tensors,
         )
 
+    # (2, 4, 4): equal dims keep a wrong axis silent (a value mismatch).
+    # (2, 3, 4): distinct dims show -1 is the last dim (a wrong axis is a reshape error).
+    @pytest.mark.parametrize(
+        "x",
+        [
+            torch.randint(-128, 127, (2, 4, 4), dtype=torch.int8),
+            torch.randint(-128, 127, (2, 3, 4), dtype=torch.int8),
+        ],
+    )
+    async def test_per_channel_negative_axis_numerical(self, x: Tensor) -> None:
+        """dequantize with a per-channel scale on a negative axis matches eager."""
+
+        class Model(nn.Module):
+            def __init__(self) -> None:
+                super().__init__()
+                self.register_buffer(
+                    "scale", torch.tensor([0.1, 0.2, 0.3, 0.4], dtype=torch.float32)
+                )
+                self.register_buffer("zero_point", torch.zeros(4, dtype=torch.int8))
+
+            def forward(self, x: Tensor) -> Tensor:
+                return torch.ops.coreai.dequantize(
+                    x, self.scale, zero_point=self.zero_point, axis=-1
+                )
+
+        model = Model()
+        await validate_numerical_output(
+            model=model, x=x, prepare_program=inject_subbyte_tensors
+        )
+
 
 # ---------------------------------------------------------------------------
 # sparse_to_dense → coreai.build_sparse_with_bitmask + coreai.sparse_with_bitmask_to_dense