Closed vidosits closed 3 years ago
If the algorithms have fundamental in-memory requirements for certain operations, you may not be able to succeed with spilling alone. It looks like you have a total of 20 GB of memory across two GPUs. Perhaps @trivialfis might have some insight into whether loading an 8GB+ sized file might push the limit with Dask XGBoost on this setup.
Wouldn't jit_unspill=True
solve that though? Wouldn't the excess just spill to RAM? Or do you mean that it can't because for some fundamental (for XGBoost) operation I'd need more than 20 GBs of GPU RAM?
The real problem was #5012. Since my data only had 1 partition there was no way to distribute it across devices. Converting to csv first and setting npartitions
solved the problem.
Happy to see it resolved!
Describe the bug I think I shouldn't be able to get OOM error with
jit_unspill=True
with a conservativedevice_memory_limit
Steps/Code to reproduce bug Input is ~8 GB CSV / 843 MB parquet.
Output with :
device_memory_limit="5GB"
or with
device_memory_limit="9GB"
Expected behavior Data spills to RAM. If
device_memory_limit
is specified and moderately low.Environment overview (please complete the following information)
Environment details
Click here to see environment details
Additional context Add any other context about the problem here.