[BUG] UMAP(n_epochs=n) transform fails for 1 and 2

[lmeyerov]

Describe the bug cuml.UMAP(n_epochs=n) fails to transform() for n=1 and n=2

Steps/Code to reproduce bug

import cudf
import cuml

df = cudf.DataFrame({'a': range(0, 100)})

for n in range(10):
    m = cuml.UMAP(n_epochs=n)
    m.fit(df)
    try:
        df2 = m.transform(df)
    except Exception as e:
        print('fail', n, e)
        continue
    print('ok', n)

=>

ok 0
fail 1 trivial_device_copy D->H failed: cudaErrorInvalidValue: invalid argument
fail 2 trivial_device_copy D->H failed: cudaErrorInvalidValue: invalid argument
ok 3
ok 4
ok 5
ok 6
ok 7
ok 8
ok 9

Expected behavior

All values in range(10) should have worked

Environment details (please complete the following information):

• python 3.10.12
• 23.10.00 in a CUDA 11.8 container (layers on top of a RAPIDS base image)
• Ubuntu 20.04 w/ CUDA 12.2 on a NVIDIA GeForce RTX 3080 Ti
• Have no tried on other systems yet

[dantegd]

Thanks for the issue @lmeyerov for the issue, I can confirm to repro on nightly 24.10 versions, not entirely sure why it's happening though...

[lmeyerov]

Yeah not too big of a deal, mostly came up as a confusing evening in why pygraphistry unit tests were failing on some new code