Handle some zero-sized corner cases in dlpack interop

[wence-]

Description

Previously, we did not explicitly check for zero-sized inputs in to/from_dlpack.
cupy>=11 changes behaviour in how it provides information for zero-sized arrays
(strides of empty tensors were previously reported as one, but now are zero).
This broke some of the validity checking in from_dlpack. To handle such cases,
check the shape of the incoming tensor as well as strides.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[wence-]

cc: @shwina, @leofang

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (branch-22.10@8ad0290). Click here to learn what that means.
Patch has no changes to coverable lines.

Additional details and impacted files

@@               Coverage Diff               @@
##             branch-22.10   #11449   +/-   ##
===============================================
  Coverage                ?   86.41%           
===============================================
  Files                   ?      145           
  Lines                   ?    22991           
  Branches                ?        0           
===============================================
  Hits                    ?    19868           
  Misses                  ?     3123           
  Partials                ?        0           

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[wence-]

This is the cudf part of the changes necessary to allow using cupy 11 and later.

[shwina]

Python looks good!

[bdice]

Thank you for the informative author comments throughout. It helped my review a lot to have a second level of description to validate my understanding of what the changes were doing. I have one minor suggestion for improving a comment.

[wence-]

I've pushed a change to the conda environment to check if running with cupy 11 actually works. I've converted to draft so this is not accidentally merged (because it might be necessary to synchronise other cupy-11-related fixes elsewhere)

[wence-]

I think everything is now good. Should we drop all the conda environment changes and do those separately?

[galipremsagar]

Looks ready to go, just need to revert the temporary changes.

[galipremsagar]

@wence- After the temporary commits are reverted and this integration PR(rapidsai/integration#528) is merged we will have to wait a couple of hours for the new env/images to be shipped so that CI can pass.

[wence-]

Note, cupy.from_dlpack which I use in favour of the deprecated cupy.fromDLpack (deprecated in cupy 10) in the tests does not exist in cupy<10. Right now cudf asks for cupy>=9.5. What's the right way to support this (requirement for backwards compat is only in test usage).

[bdice]

Note, cupy.from_dlpack which I use in favour of the deprecated cupy.fromDLpack (deprecated in cupy 10) in the tests does not exist in cupy<10. Right now cudf asks for cupy>=9.5. What's the right way to support this (requirement for backwards compat is only in test usage).

@wence- You can use packaging.version.parse to parse versions and either conditionally skip a test if the version is too low or choose the appropriate function for each range of compatible versions. Examples:

cudf/python/cudf/cudf/core/_compat.py

Lines 4 to 6 in e099e01

	from packaging import version
	PANDAS_VERSION = version.parse(pd.__version__)

cudf/python/dask_cudf/dask_cudf/io/tests/test_parquet.py

Line 321 in e099e01

if parse_version(dask.__version__) > parse_version("2021.07.0"):

[wence-]

Merged trunk to hopefully fix build errors

[wence-]

Argh, shouldn't have merged main :(.

[wence-]

@galipremsagar sorry, we both fixed the same thing and I pushed over the top of you

[galipremsagar]

@gpucibot merge