intx: pass dim through cat dispatch

The aten.cat branch of SubbyteTensor.__torch_dispatch__ reads dim via fill_defaults, then calls the op without it, so every cat runs on dim 0.
Two (2, 4) tensors with dim=1 give shape (4, 4) rather than (2, 8).
IntxTensor and UintxTensor share the branch, so both are affected.

Author: eyupcanakman

coreai_torch/_compression/_intx.py

@@ -380,7 +380,7 @@ def __torch_dispatch__(cls, func, _types, args, kwargs=None):
         if func is torch.ops.aten.cat.default:
             self, dim = fill_defaults(args, 2, [0])
             unpacked = [x.unpack_func(x.elem, x.tensor_shape, x.nbits) for x in self]
-            return cls.from_unpacked(func(unpacked), self[0].nbits)
+            return cls.from_unpacked(func(unpacked, dim), self[0].nbits)
         if func is torch.ops.aten.min.default:
             (self,) = args
             unpacked = self.unpack_func(self.elem, self.tensor_shape, self.nbits)

tests/compression/test_intx.py

@@ -85,6 +85,40 @@ def test_intx_scalar(nbits: int) -> None:
     assert intx_tensor.elem.numel() == 1
 
 
+@pytest.mark.parametrize(
+    "dim",
+    [0, 1, -1],
+)
+def test_uintx_cat(dim: int) -> None:
+    """Test uintx tensor concatenation along a given dimension."""
+    unpacked = [
+        torch.tensor([[3, 0, 14, 1], [0, 7, 15, 2]], dtype=torch.uint8),
+        torch.tensor([[1, 4, 5, 6], [2, 3, 6, 1]], dtype=torch.uint8),
+    ]
+    uintx_tensors = [UintxTensor.from_unpacked(x, 4) for x in unpacked]
+    concatenated = torch.cat(uintx_tensors, dim=dim)
+    expected = torch.cat(unpacked, dim=dim)
+    assert concatenated.shape == expected.shape
+    assert torch.equal(concatenated.to(torch.uint8), expected)
+
+
+@pytest.mark.parametrize(
+    "dim",
+    [0, 1, -1],
+)
+def test_intx_cat(dim: int) -> None:
+    """Test intx tensor concatenation along a given dimension."""
+    unpacked = [
+        torch.tensor([[3, 0, -2, 1], [0, 7, -8, 2]], dtype=torch.int8),
+        torch.tensor([[-1, 4, 5, -6], [2, -3, 6, 1]], dtype=torch.int8),
+    ]
+    intx_tensors = [IntxTensor.from_unpacked(x, 4) for x in unpacked]
+    concatenated = torch.cat(intx_tensors, dim=dim)
+    expected = torch.cat(unpacked, dim=dim)
+    assert concatenated.shape == expected.shape
+    assert torch.equal(concatenated.to(torch.int8), expected)
+
+
 def test_uint4_packed_value() -> None:
     """Test uint4 packed values."""
     unpacked = torch.tensor([3, 0, 2, 0, 9, 14, 15], dtype=torch.uint8)