fix x64 dtypes

jax/_src/interpreters/mlir.py

@@ -1003,8 +1003,6 @@ def _to_physical_op_sharding(
   if isinstance(aval, AbstractRef):
     return _to_physical_op_sharding(ctx, aval.inner_aval, sharding)
   assert isinstance(aval, (core.ShapedArray, core.DShapedArray))
-  if dtypes.issubdtype(aval.dtype, dtypes.extended):
-    assert False, aval
   axis_ctx = ctx.axis_context
   if (isinstance(axis_ctx, sharding_impls.SPMDAxisContext) and
       axis_ctx.manual_axes):
@@ -2419,9 +2417,7 @@ def wrap_with_layout_op(ctx: LoweringRuleContext,
                         aval_in: core.AbstractValue):
   result_type = aval_to_ir_type(aval_out)
   assert isinstance(result_type, ir.Type), result_type
-  if dtypes.issubdtype(aval_out.dtype, dtypes.extended):
-    assert False
-  out_shape = core.physical_aval(aval_out).shape  # type: ignore
+  out_shape = aval_out.shape  # type: ignore
   if core.is_constant_shape(out_shape):
     result_shapes = None
   else:
@@ -2659,11 +2655,8 @@ def _layout_to_mlir_layout(minor_to_major: Sequence[int] | None):
   return ir.DenseIntElementsAttr.get(layout, type=ir.IndexType.get())
 
 def _aval_to_default_layouts(aval):
-  if dtypes.issubdtype(aval.dtype, dtypes.extended):
-    assert False
-  avals = [core.physical_aval(aval)]
   # Row major order is default for `NumPy`.
-  return [list(range(aval.ndim - 1, -1, -1)) for aval in avals]
+  return [list(range(aval.ndim - 1, -1, -1))]
 
 
 def emit_python_callback(