Dry run: preprocessing::quantize::scalar - #2519
Conversation
achirkin
added
improvement
|
This PR also adds the test harness to protect newly dry run compliant code from regressions, hence most of the diff is in the tests. |
tfeher
left a comment
tfeher
left a comment
There was a problem hiding this comment.
Hi Artem, thanks for the PR, it looks good, but I have one question regarding the usage of the test util.
| template <typename Action> | ||
| void execute_with_dry_run_check(raft::resources const& res, | ||
| Action&& action, | ||
| alloc_behavior behavior, |
There was a problem hiding this comment.
What is the expected usage of this parameter?
There was a problem hiding this comment.
This harness is a verbatim copy from raft test utils. You can specify there where:
a. NO_ALLOCATIONS: you don't expect any allocations. Both dry run and real run should yield all zeroes.
b. ARGUMENT_DRIVEN: you expect the exact size of allocations is fully determined by the input arguments. Dry run and real run should return exact same sizes.
c. DATA_DRIVEN: the allocation size depends on data. Dry run sizes must never be less than real run sizes.
3 participants
Tweak scalar quantizer to support execution in dry run mode.