fix: UnboundLocalError wait in dispatcher
#4135
Conversation
| @@ -217,6 +211,10 @@ async def train_optimizer( | |||
| if batch_size > 1: # only wait if batch_size | |||
| a = self.optimizer.o_a | |||
| b = self.optimizer.o_b | |||
| wait = min( | |||
| self.max_latency * 0.95, | |||
| (batch_size * 2 + 1) * (self.a + self.optimizer.o_b), | |||
aarnphm
changed the title
wait in dispatcher
| wait = min( | ||
| self.max_latency * 0.95, | ||
| (batch_size * 2 + 1) * (a + b), | ||
| ) |
There was a problem hiding this comment.
I don't think moving wait here solves the issue
Since wait should already be defined as before
There was a problem hiding this comment.
It does since wait is only used by L219 below. But it seems to change the behavior since a, b and self.max_latency all change during the loop while wait doesn't change before.
There was a problem hiding this comment.
I agree this fix does not good solution .
It is a temporary fix Pull Request aimed at hotfix.
currently we can not use workers_per_resource Option in BentoML v1.1.0 ~v1.1.1
There was a problem hiding this comment.
@sauyon does this look ok with you?
|
This hotfix feels rushed and sloppy. thanks, I close this PR myself |
What does this PR address?
Fixes #4128
this issue is so wired
this fix won't help us figure out that why
waitdoes not exist under lineif batch_size > 1: ...I monkey-patched bentoml internal code due to this issue.