Python 3.12 xgboost.core.XGBoostError: Invalid Parameter format for nthread expect int but value='-1' when DMatrix used with import googlecloudprofiler.

[rtb-zla-karma]

Hi

I have a very peculiar error which happened when I've updated versions of Python and libs in project I'm working on.

Minimal example to reproduce the case is this:

# file.py
import googlecloudprofiler
from xgboost import DMatrix

DMatrix([[]])
print("works")

# requirements.txt
xgboost==2.0.3
google-cloud-profiler==4.1.0
#
numpy==1.26.4
scipy==1.13.0
google-api-python-client==2.125.0
google-auth==2.29.0
google-auth-httplib2==0.2.0
protobuf==4.25.3
requests==2.31.0
#
cachetools==5.3.3
certifi==2024.2.2
charset-normalizer==3.3.2
google-api-core==2.18.0
httplib2==0.22.0
idna==3.6
pyasn1==0.6.0
pyasn1_modules==0.4.0
pyparsing==3.1.2
rsa==4.9
uritemplate==4.1.1
urllib3==2.2.1

Python 3.12.2

Install with

pip install -r requirements.txt --no-deps

Run with

python file.py

Results in

Traceback (most recent call last):
  File "/project/path/file.py", line 4, in <module>
    DMatrix([[]])
  File "/venv/path/lib/python3.12/site-packages/xgboost/core.py", line 730, in inner_f
    return func(**kwargs)
           ^^^^^^^^^^^^^^
  File "/venv/path/lib/python3.12/site-packages/xgboost/core.py", line 857, in __init__
    handle, feature_names, feature_types = dispatch_data_backend(
                                           ^^^^^^^^^^^^^^^^^^^^^^
  File "/venv/path/lib/python3.12/site-packages/xgboost/data.py", line 1081, in dispatch_data_backend
    return _from_list(data, missing, threads, feature_names, feature_types)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/venv/path/lib/python3.12/site-packages/xgboost/data.py", line 1011, in _from_list
    return _from_numpy_array(array, missing, n_threads, feature_names, feature_types)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/venv/path/lib/python3.12/site-packages/xgboost/data.py", line 207, in _from_numpy_array
    _check_call(
  File "/venv/path/lib/python3.12/site-packages/xgboost/core.py", line 282, in _check_call
    raise XGBoostError(py_str(_LIB.XGBGetLastError()))
xgboost.core.XGBoostError: Invalid Parameter format for nthread expect int but value='-1'

To "solve" the problem remove import googlecloudprofiler from file.py. I really have no idea why just importing the lib causes this problem; it would make more sense after googlecloudprofiler.start is called.

I'm also going to check and post if downgrading any of main requirements, which are xgboost and google-cloud-profiler, will fix this issue. Not today though; tired.

[trivialfis]

It's probably Python 3.12 with google profiler. I looked into it a little bit, loading the _profiler.cpython-312-x86_64-linux-gnu.so inside google profiler extension causes the error. For me, it's actually a segfault during the construction of DMatrix due to an invalid write inside C++ standard function:

    #0 0x7e270a2bb8f4 in std::codecvt<wchar_t, char, __mbstate_t>::do_unshift(__mbstate_t&, char*, char*, char*&) const (/lib/x86_64-linux-gnu/libstdc++.so.6+0xbb8f4)

My guess is that it's compiled with an incompatible compiler, which can be resolved by a source installation. I don't think I can go deeper than this at the moment.

[hcho3]

It would be interesting to see if the error persists after installing the identical set of packages in a Conda environment. (Conda packages use the same C++ compiler)

[trivialfis]

the profiler is not on forge yet. Otherwise I would like to do a quick test.

[rtb-zla-karma]

Hi again.

I've tested for older lib versions we have used in Python 3.10. I have run:

pip install xgboost==1.7.4 google-cloud-profiler==4.1.0

I couldn't downgrade to google-cloud-profiler==4.0.0 because it failed to build for Python 3.12.

Anyway, code surprisingly worked with xgboost==1.7.4. Output from pip freeze:

cachetools==5.3.3
certifi==2024.2.2
charset-normalizer==3.3.2
google-api-core==2.18.0
google-api-python-client==2.127.0
google-auth==2.29.0
google-auth-httplib2==0.2.0
google-cloud-profiler==4.1.0
googleapis-common-protos==1.63.0
httplib2==0.22.0
idna==3.7
numpy==1.26.4
proto-plus==1.23.0
protobuf==4.25.3
pyasn1==0.6.0
pyasn1_modules==0.4.0
pyparsing==3.1.2
requests==2.31.0
rsa==4.9
scipy==1.13.0
uritemplate==4.1.1
urllib3==2.2.1
xgboost==1.7.4

I've also tested it for in-between versions of xgboost and the highest version the code worked with is 1.7.6 and it started failing since 2.0.0.

[trivialfis]

I couldn't downgrade to google-cloud-profiler==4.0.0 because it failed to build for Python 3.12.

I suggest opening an issue for google cloud profiler to fix this first. ;-)

[rtb-zla-karma]

It's still bad TBH. After downgrading I get another error later in the project code when initialized DMatrix is used for predictions using model. I won't paste the code just the exception from minimal example:

Traceback (most recent call last):
  File "/project/path/file2.py", line 19, in <module>
    main()
  File "/project/path/file2.py", line 16, in main
    model.predict(matrix)[0].item()
    ^^^^^^^^^^^^^^^^^^^^^
  File "/venv/path/lib/python3.12/site-packages/xgboost/core.py", line 2163, in predict
    _check_call(
  File "/venv/path/lib/python3.12/site-packages/xgboost/core.py", line 279, in _check_call
    raise XGBoostError(py_str(_LIB.XGBGetLastError()))
xgboost.core.XGBoostError: Invalid Parameter format for disable_default_eval_metric expect boolean but value=''

Again, the solution is to remove import googlecloudprofiler. It seems that is have to remove this lib altogether.