pretrain阶段，streaming 和 packing不兼容的问题

Checklist / 检查清单

• I have searched existing issues, and this is a new bug report. / 我已经搜索过现有的 issues，确认这是一个新的 bug report。

Bug Description / Bug 描述

Rank(Worker): 0 caught user code error, which triggered following unhandled exception:
Exception type: TypeError
Detail:
Traceback (most recent call last):
File "/checkpoint/binary/train_package/swift/cli/pt.py", line 7, in
pretrain_main()
File "/checkpoint/binary/train_package/swift/pipelines/train/pretrain.py", line 17, in pretrain_main
return SwiftPretrain(args).main()
File "/checkpoint/binary/train_package/swift/pipelines/base.py", line 52, in main
result = self.run()
File "/checkpoint/binary/train_package/swift/ray/base.py", line 168, in wrapper
return func(self, *args, **kwargs)
File "/checkpoint/binary/train_package/swift/pipelines/train/sft.py", line 174, in run
train_dataset, val_dataset = self._prepare_dataset()
File "/checkpoint/binary/train_package/swift/ray/base.py", line 168, in wrapper
return func(self, *args, **kwargs)
File "/checkpoint/binary/train_package/swift/pipelines/train/sft.py", line 135, in _prepare_dataset
self._show_dataset(*datasets)
File "/checkpoint/binary/train_package/swift/pipelines/train/sft.py", line 295, in _show_dataset
inputs = train_dataset[0] if hasattr(train_dataset, 'len') else next(iter(train_dataset))
File "/checkpoint/binary/train_package/swift/dataset/packing.py", line 196, in iter
data = self._fetch_data_out_queue(data, num_samples)
File "/checkpoint/binary/train_package/swift/dataset/packing.py", line 171, in _fetch_data_out_queue
res[i] = (data, len(data['input_ids']))
TypeError: list indices must be integers or slices, not str

让llm查了一下，发现是packing的时候需要input id，但是streaming的时候做不到这点，故而报错。

How to Reproduce / 如何复现

args="--model ${CKPT_PATH}
--max_steps 5000
--save_strategy steps
--shuffle_buffer_size 10000
--streaming true
--model_type qwen3
--train_type full
--dataset ${DATA_PATHS[@]}
--val_dataset ${VAL_DATA_PATHS[@]}
--packing true
--packing_num_proc ${num}
--enable_channel_loss true
--per_device_eval_batch_size 2
--eval_steps 50
--dataset_shuffle true
--torch_dtype bfloat16
--per_device_train_batch_size 2
--sequence_parallel_size 1
--learning_rate 1e-5
--gradient_accumulation_steps 2
--num_train_epochs 1
--truncation_strategy split
--max_length 8192
--save_steps 200
--logging_steps 1
--save_total_limit 100
--warmup_ratio 0.05
--dataloader_num_workers ${num}
--dataset_num_proc ${num}
--save_only_model true
--overwrite_output_dir true
--deepspeed zero3
--attn_impl flash_attention_2
--output_dir /data/oss_bucket_0/qingxi_hzx/cpt_model/${SCRIPT_NAME}/
--report_to swanlab
--run_name ${SCRIPT_NAME}
--swanlab_token ${SWANLAB_API_KEY}
--swanlab_project ${SWANLAB_PROJECT}"

Additional Information / 补充信息

No response

Start with dataset/packing.py, especially iter and _fetch_data_out_queue, then inspect the dataset preparation and preview flow in pipelines/train/sft.py. Reproduce with streaming true and packing true using the provided pretraining arguments. Done means this combination no longer raises the reported TypeError and the dataset preview or training can proceed.