【Bug】RuntimeError: split_with_sizes():not supported for NestedTensor on dim=0

[hustmf]

System Info

verl version: main ,commit id f3a0233
platform: A100
python:3.12
torch:2.8.0
主线engine_worker实现bug, 在训练的前向切分microbatch时, 不开启dynamic batch size情况下，直接使用tensordict的split方法不支持NestedTensor
代码实现位置
报错堆栈：
(TaskRunner pid=1953021) File "verl/verl/workers/engine_workers.py", line 323, in infer_batch
(TaskRunner pid=1953021) output = self.engine.infer_batch(data, loss_function=loss_function)
(TaskRunner pid=1953021) File "verl/verl/workers/engine/base.py", line 140, in infer_batch
(TaskRunner pid=1953021) outputs = self.forward_backward_batch(data, loss_function, forward_only=True)
(TaskRunner pid=1953021) File "verl/verl/workers/engine/fsdp/transformer_impl.py", line 499, in forward_backward_batch
(TaskRunner pid=1953021) micro_batches, indices = prepare_micro_batches(
(TaskRunner pid=1953021) File "verl/verl/workers/engine/utils.py", line 88, in prepare_micro_batches
(TaskRunner pid=1953021) micro_batches = data.split(micro_batch_size_per_gpu)
(TaskRunner pid=1953021) File "/root/miniconda3/envs/verl-sgl-mf-py312/lib/python3.12/site-packages/tensordict/_td.py", line 1774, in split
(TaskRunner pid=1953021) splits = {k: v.split(splits, dim) for k, v in self.items()}
(TaskRunner pid=1953021) File "/root/miniconda3/envs/verl-sgl-mf-py312/lib/python3.12/site-packages/torch/nested/_internal/ops.py", line 54, in _wrap_jagged_dim
(TaskRunner pid=1953021) raise RuntimeError(f"{op_name}(): not supported for NestedTensor on dim=0")
(TaskRunner pid=1953021) RuntimeError: split_with_sizes(): not supported for NestedTensor on dim=0
初步定位分析：
原有fsdpworker、megatronworker实现中切分batch时数据类型为DataProto，自己定义了split方法，最新engine worker中切分时数据类型直接为TensorDict，batch中的input_ids和position_ids为NestedTensor，形状不规则

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

grpo主线脚本
use_legacy_worker_impl=disable
use_dynamic_bsz=False

Expected behavior

计算old log p时前向切分batch报错