为什么在 WSL 中使用 RAPIDS 时会出现 CUDA 内存错误？

Question

我在 Windows 21H2 (19044.1348) 下安装了 WSL 2 (5.10.60.1-microsoft-standard-WSL2)，并使用带有 Pascal GPU (1070) 的 NVidia 驱动程序 510.06。我在 WSL (20.04.3 LTS) 中使用默认的 ubuntu 版本我尝试了 docker 和 anaconda 版本。我可以运行 Jupiter Notebook 并导入图书馆的。您还可以创建一个 cudf 数据报。但是写入它或发出任何其他内容都会导致内存错误。

buf = rmm.DeviceBuffer(size=100)

给我（有一次运行没有错误，但现在没有了）

---------------------------------------------------------------------------
MemoryError                               Traceback (most recent call last)
/tmp/ipykernel_2220/3317065296.py in <module>
      1 import rmm
----> 2 buf = rmm.DeviceBuffer(size=100)

rmm/_lib/device_buffer.pyx in rmm._lib.device_buffer.DeviceBuffer.__cinit__()

MemoryError: std::bad_alloc: CUDA error at: /home/user/miniconda3/envs/rapids-21.10/include/rmm/mr/device/cuda_memory_resource.hpp:70: cudaErrorNotSupported operation not supported

和

gdf_float = cudf.DataFrame()
gdf_float['0'] = [1.0, 2.0, 5.0]
gdf_float['1'] = [4.0, 2.0, 1.0]
gdf_float['2'] = [4.0, 2.0, 1.0]

给我

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
~/miniconda3/envs/rapids-21.10/lib/python3.7/site-packages/cudf/core/column/column.py in as_column(arbitrary, nan_as_null, dtype, length)
   2026             data = as_column(
-> 2027                 memoryview(arbitrary), dtype=dtype, nan_as_null=nan_as_null
   2028             )

TypeError: memoryview: a bytes-like object is required, not 'list'

During handling of the above exception, another exception occurred:

RuntimeError                              Traceback (most recent call last)
/tmp/ipykernel_2220/2068985133.py in <module>
      1 gdf_float = cudf.DataFrame()
----> 2 gdf_float['0'] = [1.0, 2.0, 5.0]
      3 gdf_float['1'] = [4.0, 2.0, 1.0]
      4 gdf_float['2'] = [4.0, 2.0, 1.0]

~/miniconda3/envs/rapids-21.10/lib/python3.7/contextlib.py in inner(*args, **kwds)
     72         def inner(*args, **kwds):
     73             with self._recreate_cm():
---> 74                 return func(*args, **kwds)
     75         return inner
     76 

~/miniconda3/envs/rapids-21.10/lib/python3.7/site-packages/cudf/core/dataframe.py in __setitem__(self, arg, value)
    766                     # disc. with pandas here
    767                     # pandas raises key error here
--> 768                     self.insert(len(self._data), arg, value)
    769 
    770         elif can_convert_to_column(arg):

~/miniconda3/envs/rapids-21.10/lib/python3.7/contextlib.py in inner(*args, **kwds)
     72         def inner(*args, **kwds):
     73             with self._recreate_cm():
---> 74                 return func(*args, **kwds)
     75         return inner
     76 

~/miniconda3/envs/rapids-21.10/lib/python3.7/site-packages/cudf/core/dataframe.py in insert(self, loc, name, value)
   3276             )
   3277 
-> 3278         value = column.as_column(value)
   3279 
   3280         self._data.insert(name, value, loc=loc)

~/miniconda3/envs/rapids-21.10/lib/python3.7/site-packages/cudf/core/column/column.py in as_column(arbitrary, nan_as_null, dtype, length)
   2100                     ),
   2101                     dtype=dtype,
-> 2102                     nan_as_null=nan_as_null,
   2103                 )
   2104             except (pa.ArrowInvalid, pa.ArrowTypeError, TypeError):

~/miniconda3/envs/rapids-21.10/lib/python3.7/site-packages/cudf/core/column/column.py in as_column(arbitrary, nan_as_null, dtype, length)
   1794                 "https://issues.apache.org/jira/browse/ARROW-3802"
   1795             )
-> 1796         col = ColumnBase.from_arrow(arbitrary)
   1797         if isinstance(arbitrary, pa.NullArray):
   1798             if type(dtype) == str and dtype == "empty":

~/miniconda3/envs/rapids-21.10/lib/python3.7/site-packages/cudf/core/column/column.py in from_arrow(cls, array)
    305             return cudf.core.column.Decimal64Column.from_arrow(array)
    306 
--> 307         result = libcudf.interop.from_arrow(data, data.column_names)[0]["None"]
    308 
    309         result = result._with_type_metadata(

cudf/_lib/interop.pyx in cudf._lib.interop.from_arrow()

RuntimeError: CUDA error encountered at: ../src/bitmask/null_mask.cu:93: 801 cudaErrorNotSupported operation not supported

如果这是相关的我的系统内存总是接近满 (16GB)，Vmmem 占用大约 10GB 我的显存只有 1,4/8GB

Answer 1

遗憾的是，WSL2 上的 RAPIDS 只能在具有 RAPIDS 21.08 的 Pascal GPU 上运行，而不能在 21.10 或更高版本上运行。请尝试 21.08。这些版本仍处于试验阶段，因此 YMMV。

为什么在 WSL 中使用 RAPIDS 时会出现 CUDA 内存错误？

Why do I get a CUDA memory error when using RAPIDS in WSL?

ubuntu

cuda

windows-subsystem-for-linux

rapids

cudf