Move vectorized domains innermost always

[naoyam]

pytorch/third_party/nvfuser/csrc/scheduler/reduction_utils.cpp

Lines 259 to 276 in 7b37a83

	// In the case of outer grid persistence, make sure the vectorized
	// domain placed at the innermost position.
	// TODO: Why isn't this the case by default?
	if (is_outer_grid_persistence) {
	int vec_id_cur_pos = -1;
	std::unordered_map<int, int> vec_reorder_map;
	for (const auto i : c10::irange(reduction_rf_tv->nDims())) {
	auto id = reduction_rf_tv->axis(i);
	if (id->getParallelType() == ParallelType::Vectorize) {
	vec_id_cur_pos = i;
	vec_reorder_map[i] = -1;
	} else if (vec_id_cur_pos >= 0) {
	vec_reorder_map[i] = i - 1;
	}
	}
	TORCH_INTERNAL_ASSERT(vec_id_cur_pos != -1, "Vectorized ID not found");
	reduction_rf_tv->reorder(vec_reorder_map);
	}