Handle invalid BroadcastTo shape in C backend #175
Conversation
ricardoV94
changed the title
ricardoV94
force-pushed
the
broadcast_to_check
branch
from
555f2a4 to
43a0a79
Compare
ricardoV94
changed the title
ricardoV94
force-pushed
the
broadcast_to_check
branch
from
43a0a79 to
df568d9
Compare
ricardoV94
force-pushed
the
broadcast_to_check
branch
2 times, most recently
from
8be1bf8 to
3c43f0a
Compare
tests/tensor/test_extra_ops.py
Outdated
| @@ -1313,7 +1333,7 @@ def test_memory_leak(self): | |||
| blocks_last = blocks_i | |||
|
|
|||
| tracemalloc.stop() | |||
| assert np.allclose(np.mean(block_diffs), 0) | |||
| assert np.mean(block_diffs) <= (0 + 1e-8) | |||
There was a problem hiding this comment.
This was failing with negative values, which should not be problematic (?). No idea what the function does when it raises.
ricardoV94
force-pushed
the
broadcast_to_check
branch
from
3c43f0a to
883c963
Compare
There was a problem hiding this comment.
LGTM @ricardoV94. I don't know what do make out of the performance regression. I doesn't seem related to this PR but I can't seem to find a way to get the detailed benchmark results from the cache or from this PR.
|
The last cached benchmark might have been a CI fluke and this one looks bad in comparison? Only the C scan test could possibly have been affected so I'll run it locally before and after the changes to see if it complains. |
|
Looks like the performance regression problems are happening because the tests that use random are not being seeded, and different seeds can lead to very different run times |
Closes #152