DataLoader2 with multiprocess raise exception: Can not request next item while we are still waiting response for previous request

[npuichigo]

🐛 Describe the bug

dp = ...
dp = dp.sharding_filter()
rs = MultiProcessingReadingService(num_workers=4)
dataloader = DataLoader2(dp, reading_service=rs)

for _ in dataloader:
    pass

dataloader.shutdown()

Exception: Can not request next item while we are still waiting response for previous request
This exception is thrown by __iter__ of _IterateQueueDataPipes(datapipes=[QueueWrapper, QueueWrapper, QueueWrapper, QueueWrapper])

Versions

[pip3] numpy==1.26.4
[pip3] onnx==1.15.0
[pip3] onnxconverter-common==1.13.0
[pip3] onnxruntime==1.15.1
[pip3] skl2onnx==1.16.0
[pip3] torch==2.2.1
[pip3] torchaudio==2.2.1
[pip3] torchdata==0.7.1
[pip3] torchvision==0.17.1
[pip3] triton==2.2.0
[conda] numpy                     1.26.4                   pypi_0    pypi
[conda] torch                     2.2.1                    pypi_0    pypi
[conda] torchaudio                2.2.1                    pypi_0    pypi
[conda] torchvision               0.17.1                   pypi_0    pypi
[conda] triton

[jdenhof]

I have noticed this happens whenever the number of workers you specify for the MultiProcessingReadingService is greater than then the number elements that can be yielded from the dp before sharding.