[REVIEW] BUG Remove deprecated call to from_gpu_matrix

[hlinsen]

This PR removes deprecated method DataFrame.from_gpu_matrix
Fixes: #1191

[GPUtester]

Please update the changelog in order to start CI tests.

View the gpuCI docs here.

[afender]

👍

[afender]

I've seen this test error on PR #1156 as well

[BradReesWork]

rerun tests

[afender]

@raydouglass was there a recent change on the CI side that could cause all cugraph python tests to segfault in cudf (+ @kkraus14)?

Looking at cugraph merge history and cudf merge history of the last day I don't really see anything suspicious.

[kkraus14]

@raydouglass was there a recent change on the CI side that could cause all cugraph python tests to segfault in cudf (+ @kkraus14)?

Looks to be CSV related:

 4  /opt/conda/envs/rapids/lib/python3.8/site-packages/cudf/_lib/csv.cpython-38-x86_64-linux-gnu.so(_ZNSt6vectorIN4cudf2io16column_name_infoESaIS2_EED1Ev+0xc4) [0x7f0901d69234]
 5  /opt/conda/envs/rapids/lib/python3.8/site-packages/cudf/_lib/csv.cpython-38-x86_64-linux-gnu.so(_ZN4cudf2io19table_with_metadataD1Ev+0xd6) [0x7f0901d6b0e6]
 6  /opt/conda/envs/rapids/lib/python3.8/site-packages/cudf/_lib/csv.cpython-38-x86_64-linux-gnu.so(+0x45460) [0x7f0901d51460]
 7  /opt/conda/envs/rapids/lib/python3.8/site-packages/cudf/_lib/csv.cpython-38-x86_64-linux-gnu.so(+0x4f657) [0x7f0901d5b657]

I know these APIs were under some churn recently, cc @vuule

[kkraus14]

Also, looks like the libcudf and cudf builds that were build were mismatched which is a possible issue.

Would suggest adding libcudf / librmm here as well: https://github.com/rapidsai/cugraph/blob/branch-0.16/ci/gpu/build.sh#L60

This should help the conda solver prioritize getting the newest version of libcudf / librmm.

[afender]

Quick updates:

I added the change to ci/gpu/build.sh#L60 in #1199 and still get the version mismatch :

cudf-0.16.0a201007         |cuda_11.0_py38_g8950c1a69e_1962        86.3 MB  rapidsai-nightly
libcudf-0.16.0a201006      |cuda11.0_g953f825b74_1958       242.5 MB  rapidsai-nightly

I did check cuML and they have the same mismatch. However, CI passes there.

[kkraus14]

Quick updates:

I added the change to ci/gpu/build.sh#L60 in #1199 and still get the version mismatch :

cudf-0.16.0a201007         |cuda_11.0_py38_g8950c1a69e_1962        86.3 MB  rapidsai-nightly
libcudf-0.16.0a201006      |cuda11.0_g953f825b74_1958       242.5 MB  rapidsai-nightly

I did check cuML and they have the same mismatch. However, CI passes there.

They're likely not using the read_csv function which would have mismatched symbols. I'm guessing this is from us removing the dlpack dependency and subsequently adding it back in. I'm not sure how to fix this other than changing the way we pin the libcudf dependency in cudf to the exact build.

[BradReesWork]

rerun tests

[BradReesWork]

rerun tests

[BradReesWork]

rerun tests

[BradReesWork]

rerun tests

[rlratzel]

rerun tests

[codecov-io]

Codecov Report

Merging #1198 into branch-0.16 will not change coverage.
The diff coverage is n/a.

@@             Coverage Diff              @@
##           branch-0.16    #1198   +/-   ##
============================================
  Coverage        57.28%   57.28%           
============================================
  Files               61       61           
  Lines             2500     2500           
============================================
  Hits              1432     1432           
  Misses            1068     1068           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update b14c458...f608570. Read the comment docs.