cudf groupby udf & observed

python/xorbits/_mars/dataframe/groupby/tests/test_groupby_execution.py

@@ -1079,7 +1079,6 @@ def __call__(self, s):
  )
 
 
-@support_cuda
 @pytest.mark.parametrize(
  "chunked,as_index", [(True, True), (True, False), (False, True), (False, False)]
 )
@@ -1092,6 +1091,7 @@ def test_groupby_apply_as_index(chunked, as_index, setup_gpu, gpu):
  }
  )
 
+ # cudf not support udf like this
  def udf(v):
  denominator = v["a"].sum() * v["a"].mean()
  v = v[v["c"] == "c"]

python/xorbits/_mars/dataframe/reduction/core.py

@@ -1161,9 +1161,7 @@ def _interpret_var(v):
  and isinstance(t.op.lhs, DATAFRAME_TYPE)
  and isinstance(t.op.rhs, str)
  ):
- # for a cudf dataframe, df == 'foo' doesn't work, so we convert the rhs
- # to a tuple.
- rhs = f"({rhs},) * len({lhs}.columns)"
+ rhs = f"{rhs} "
  statements = [
  f"try:",
  f" {var_name} = {lhs}.{func_name}({rhs}, {axis_expr})",

python/xorbits/_mars/lib/groupby_wrapper.py

@@ -290,6 +290,10 @@ def wrapped_groupby(
  groupby_kw.pop("squeeze")
  if not _HAS_DROPNA: # pragma: no branch
  groupby_kw.pop("dropna")
+ # cudf currently not support observed,
+ # cudf default to True, while pandas default is False
+ if is_cudf(obj): # pragma: no branch
+ groupby_kw["observed"] = True
 
  groupby_obj = obj.groupby(**groupby_kw)
  return GroupByWrapper(obj, groupby_obj=groupby_obj, as_index=as_index)