ERNIE-4.5-VL-28B-A3B-Paddle sft 训练报错：AssertionError: weight_idxs' length should be greater than 0

[lx11]

根据官方环境：
paddle_xpu 0.0.1
paddle2onnx 2.0.1
paddleformers 0.4.0
paddlepaddle-xpu 3.3.0.dev20251016

启动命令：erniekit train examples/configs/xpu/ERNIE-4.5-VL-28B-A3B/sft/run_sft_32k.yaml

[2025-12-23 12:07:45,066] [ INFO] pp_layers.py:854 - start segment network..
Traceback (most recent call last):
File "/work/ERNIE/erniekit/launcher.py", line 58, in
launch()
File "/work/ERNIE/erniekit/launcher.py", line 46, in launch
run_tuner()
File "/work/ERNIE/erniekit/train/tuner.py", line 82, in run_tuner
_training_function(config={"args": args})
File "/work/ERNIE/erniekit/train/tuner.py", line 59, in _training_function
run_vl_sft(
File "/work/ERNIE/erniekit/train/vl_sft/workflow.py", line 517, in run_vl_sft
model = Ernie4_5_VLMoeForConditionalGenerationPipe.from_pretrained(
File "/usr/local/lib/python3.10/dist-packages/paddleformers/transformers/model_utils.py", line 2965, in from_pretrained
model = cls(config, *init_args, **model_kwargs)
File "/usr/local/lib/python3.10/dist-packages/paddleformers/transformers/utils.py", line 290, in impl
init_func(self, *args, **kwargs)
File "/work/ERNIE/ernie/modeling_moe_vl_pp.py", line 1908, in init
PipelineLayer.init(
File "/usr/local/lib/python3.10/dist-packages/paddle/distributed/fleet/meta_parallel/parallel_layers/pp_layers.py", line 506, in init
self._segment_network(seg_method)
File "/usr/local/lib/python3.10/dist-packages/paddle/distributed/fleet/meta_parallel/parallel_layers/pp_layers.py", line 858, in _segment_network
self.segment_parts = seg.do_segment()
File "/usr/local/lib/python3.10/dist-packages/paddle/distributed/fleet/meta_parallel/parallel_layers/pp_layers.py", line 187, in do_segment
weight_idxs = self._gen_layer_weight(layername)
File "/usr/local/lib/python3.10/dist-packages/paddle/distributed/fleet/meta_parallel/parallel_layers/pp_layers.py", line 234, in _gen_layer_weight
assert len(weight_idxs) > 0, (
AssertionError: weight_idxs' length should be greater than 0