[diffusion] benchmark: enable performance benchmark with warmup#15773
[diffusion] benchmark: enable performance benchmark with warmup#15773fsygd wants to merge 5 commits intosgl-project:mainfrom
Conversation
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
| return | ||
|
|
||
| bootstrap_info = f", bootstrap_room={self.bootstrap_room}" if self.bootstrap_room is not None else "" | ||
| bootstrap_info = ( |
There was a problem hiding this comment.
Pre-commit did this, not me. Should I rollback?
|
@BBuf would you take a look? |
|
Any blocker? @mickqian |
|
If warmup is enabled by default here, it might mask some performance issues that would otherwise be discoverable, such as the problem found previously here: #15511. For offline inference, you must accept the warmup cost. For serving, the first request also has to accept the warmup cost. For subsequent requests, if the resolution changes or similar things happen, deepgemm still needs to recompile, which also has a cost. These should not be removed, especially for profiling. |
I see, warmup by default may hide something we concern. May i add warmup for option maybe sometimes we need it, or just close the pr? |
51a933c to
876f881
Compare
876f881 to
fa6535d
Compare
|
Warmup is an option now, PTAL @mickqian @BBuf @RubiaCx @FlamingoPg |
Motivation
Performance benchmark will be better with warmup to get more accurate performance, especially when something like DeepGemm is involved. (e.g. #15403)
without warmup on H200:
with warmup on H100:
Modifications
Accuracy Tests
Benchmarking and Profiling
Checklist