rapidsai/cudf
[BUG] `cudf.read_text` throws an exception when reading a host buffer
Closed
#13,734 opened on Jul 22, 2023
0 - BacklogPythonbugcuIOgood first issue
Repository metrics
- Stars
- (6,000 stars)
- PR merge metrics
- (Avg merge 17d 21h) (230 merged PRs in 30d)
Description
Describe the bug
cudf.read_text throws an exception when reading a host buffer, unlike the CSV, Parquet, ORC and JSON readers. Hopefully it is straightforward to enable cudf.read_text to read from host buffers.
Steps/Code to reproduce bug Here is a small code snippet that shows the exception:
from io import BytesIO
import cudf
df = cudf.DataFrame({'a':['aaaa','bbbb']})
buf = BytesIO()
df.to_csv(buf, index=False)
df2 = cudf.read_csv(buf) # this is ok
print(df2)
buf = BytesIO()
df.to_csv(buf, index=False)
df2 = cudf.read_text(buf, delimiter='\n') # this crashes
print(df2)
Expected behavior
I expect that read_text would support host buffers correctly in the python layer. "multibyte_split" works fine with HOST_BUFFER data source in the libcudf benchmarks.
Environment overview (please complete the following information) nightly docker image from 23.08, A100 DGX workstation