Failed running demo/guide-python/external_memory.py built from main

[SkqLiao]

Hi, I'm trying to use ExtMemQuantileDMatrix for training huge dateset on gpus.

For example, training 1Tb raw fp32 dataset on 4/8xRTX 4090(24G) + 2/4Tb memory(which is sufficent for the same dataset with CPU). (btw, is it possible?)

However, I encountered this error when running the demo code demo/guide-python/external_memory.py:

/cache/xgboost/python-package/xgboost/core.py:1893: UserWarning: [01:54:44] WARNING: /cache/xgboost/src/data/ellpack_page_source.h:191: CUDA heterogeneous memory management is not available. The overhead of iterating through external memory might be significant.
  self._init(
/cache/xgboost/python-package/xgboost/core.py:1893: UserWarning: [01:54:44] WARNING: /cache/xgboost/src/data/ellpack_page_source.cu:618: Running on a NUMA system without membind. The overhead of iterating through external memory might be significant.
  self._init(
Traceback (most recent call last):
  File "/cache/xgboost/demo/guide-python/external_memory.py", line 213, in <module>
    main(tmpdir, args)
  File "/cache/xgboost/demo/guide-python/external_memory.py", line 172, in main
    hist_train(it)
  File "/cache/xgboost/demo/guide-python/external_memory.py", line 136, in hist_train
    Xy = xgboost.ExtMemQuantileDMatrix(it, missing=np.nan, enable_categorical=False)
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/cache/xgboost/python-package/xgboost/core.py", line 774, in inner_f
    return func(**kwargs)
           ^^^^^^^^^^^^^^
  File "/cache/xgboost/python-package/xgboost/core.py", line 1893, in __init__
    self._init(
  File "/cache/xgboost/python-package/xgboost/core.py", line 1940, in _init
    _check_call(ret)
  File "/cache/xgboost/python-package/xgboost/core.py", line 323, in _check_call
    raise XGBoostError(py_str(_LIB.XGBGetLastError()))
xgboost.core.XGBoostError: [01:54:44] /cache/xgboost/src/common/common.cu:16: /cache/xgboost/src/common/cuda_pinned_allocator.cu: 49: cudaErrorInvalidValue: invalid argument
Stack trace:
  [bt] (0) /cache/xgboost/python-package/xgboost/../../lib/libxgboost.so(+0x4c8bb1) [0x7f0fa1bc1bb1]
  [bt] (1) /cache/xgboost/python-package/xgboost/../../lib/libxgboost.so(dh::ThrowOnCudaError(cudaError, char const*, int)+0x603) [0x7f0fa24ee453]
  [bt] (2) /cache/xgboost/python-package/xgboost/../../lib/libxgboost.so(+0xdf5900) [0x7f0fa24ee900]
  [bt] (3) /cache/xgboost/python-package/xgboost/../../lib/libxgboost.so(xgboost::common::cuda_impl::CreateHostMemPool()+0x9) [0x7f0fa24ee969]
  [bt] (4) /cache/xgboost/python-package/xgboost/../../lib/libxgboost.so(xgboost::data::EllpackMemCache::EllpackMemCache(xgboost::data::EllpackCacheInfo, int)+0x306) [0x7f0fa260dd66]
  [bt] (5) /cache/xgboost/python-package/xgboost/../../lib/libxgboost.so(xgboost::data::EllpackCacheStreamPolicy<xgboost::EllpackPage, xgboost::data::EllpackFormatPolicy>::CreateWriter(xgboost::StringView, unsigned int)+0x3cb) [0x7f0fa26161db]
  [bt] (6) /cache/xgboost/python-package/xgboost/../../lib/libxgboost.so(xgboost::data::SparsePageSourceImpl<xgboost::EllpackPage, xgboost::data::EllpackCacheStreamPolicy<xgboost::EllpackPage, xgboost::data::EllpackFormatPolicy> >::WriteCache()+0x84) [0x7f0fa261af54]
  [bt] (7) /cache/xgboost/python-package/xgboost/../../lib/libxgboost.so(xgboost::data::ExtEllpackPageSourceImpl<xgboost::data::EllpackCacheStreamPolicy<xgboost::EllpackPage, xgboost::data::EllpackFormatPolicy> >::ExtEllpackPageSourceImpl(xgboost::Context const*, xgboost::MetaInfo*, xgboost::data::ExternalDataInfo, std::shared_ptr<xgboost::data::Cache>, std::shared_ptr<xgboost::common::HistogramCuts>, std::shared_ptr<xgboost::data::DataIterProxy<void (void*), int (void*)> >, xgboost::data::DMatrixProxy*, xgboost::data::EllpackCacheInfo const&)+0x743) [0x7f0fa262ab83]
  [bt] (8) /cache/xgboost/python-package/xgboost/../../lib/libxgboost.so(xgboost::data::ExtMemQuantileDMatrix::InitFromCUDA(xgboost::Context const*, std::shared_ptr<xgboost::data::DataIterProxy<void (void*), int (void*)> >, void*, xgboost::BatchParam const&, std::shared_ptr<xgboost::DMatrix>, long, xgboost::ExtMemConfig const&)+0xa47) [0x7f0fa26266e7]

I installed XGBoost according to the document:

git clone --recursive https://github.com/dmlc/xgboost
cd xgboost
cmake -B build -S . -DUSE_CUDA=ON -DUSE_NCCL=ON -DPLUGIN_RMM=ON -DCMAKE_PREFIX_PATH=$CONDA_PREFIX  -DBUILD_WITH_SHARED_NCCL=ON
cd build && make -j$(nproc)
cd ../python-package && pip install -e .

The enviroment is RTX 4090 with cuda-12.9 and 2Tb memory. Thanks!

[trivialfis]

Could you please share the output from nvidia-smi?

On the other hand, it's not recommended to use external memory with PCIe-based devices due to bandwidth limitations. Please visit the document https://xgboost.readthedocs.io/en/stable/tutorials/external_memory.html .

[SkqLiao]

@trivialfis Thanks — I discovered that it works fine on another 8×H100 cluster. I realized the failure was caused by a mismatch between CUDA and the driver (CUDA 12.9 was too high). I removed the nvcc installation from Conda and rebuilt XGBoost, and now it runs correctly.

However, I later found that after reading all batches, I encounter an OOM error when the dataset is large. I have set min_cache_page_bytes to 64MB and cache_host_ratio to 1 . Do you have any suggestions for further reducing GPU memory usage (I’d like to test the limit ratio between GPU memory and dataset size, ignoring performance for now)?
Or, what's the mininum gpu memory usage for ExtMemQuantileDMatrix? Thanks.

Traceback (most recent call last):
  File "/cache/gpu-training/py_src/xgb.py", line 382, in <module>
    main(args)
  File "/cache/gpu-training/py_src/xgb.py", line 263, in main
    hist_train(
  File "/cache/gpu-training/py_src/xgb.py", line 139, in hist_train
    Xy = xgboost.ExtMemQuantileDMatrix(
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/cache/xgboost/python-package/xgboost/core.py", line 774, in inner_f
    return func(**kwargs)
           ^^^^^^^^^^^^^^
  File "/cache/xgboost/python-package/xgboost/core.py", line 1893, in __init__
    self._init(
  File "/cache/xgboost/python-package/xgboost/core.py", line 1940, in _init
    _check_call(ret)
  File "/cache/xgboost/python-package/xgboost/core.py", line 323, in _check_call
    raise XGBoostError(py_str(_LIB.XGBGetLastError()))
xgboost.core.XGBoostError: [12:13:13] /cache/xgboost/src/common/device_vector.cu:23: Memory allocation error on worker 0: std::bad_alloc: out_of_memory: RMM failure at:/tmp/pip-build-env-n55gyfyv/normal/lib/python3.11/site-packages/librmm/include/rmm/mr/device/arena_memory_resource.hpp:171: Maximum pool size exceeded (failed to allocate 4.026115 GiB): No room in arena.
- Free memory: 1.90009GB
- Requested memory: 4.02612GB