search

nv-legate / cupynumeric

An Aspiring Drop-In Replacement for NumPy at Scale
https://docs.nvidia.com/cupynumeric
Apache License 2.0
623 stars 71 forks source link

Update cuda version to 12.2 #1120

Closed mag1cp1n closed 9 months ago

mag1cp1n commented 9 months ago

Also bump the legate.core version to TOT.

aschaffer commented 9 months ago

I've managed to repro the test_overlap.py failures, locally. If allclose() (with eps=1.e-8) is used then the tests PASS, at least locally on my node.

from utils.comparisons import allclose
#...
    eps = 1.0e-8
    # 
    # assert np.array_equal(a_np, a_num)
    #
    # this fixes the failure:
    #
    assert allclose(a_np, a_num, atol=eps)
    #...
manopapad commented 9 months ago

Thank you @aschaffer. This was my mistake. I should have used allclose in that testcase in the first place, since it was comparing the results of floating point computations from two different implementations. Weirdly this didn't show up originally, but apparently the CUDA bump caused the result of sin to change ever so slightly.

aschaffer commented 9 months ago

Thank you @aschaffer. This was my mistake. I should have used allclose in that testcase in the first place, since it was comparing the results of floating point computations from two different implementations. Weirdly this didn't show up originally, but apparently the CUDA bump caused the result of sin to change ever so slightly.

No worries. Yup, that typically happens: CTK, or other library dependency changes and round-off errors change ever so slightly, but enough to fail "exact" comparisons.