PERF: delay frontends initialization until yt.load is actually called #4539
Conversation
neutrinoceros
force-pushed
the
delay_frontends_load
branch
5 times, most recently
from
ed268cd
to
f904af2
Compare
neutrinoceros
force-pushed
the
delay_frontends_load
branch
from
f904af2
to
e565cd6
Compare
yt/frontends/__init__.py
Outdated
| if value == "_all": | ||
| for _ in __all__: | ||
| __getattr__(_) | ||
| return |
There was a problem hiding this comment.
note: the reason this special case is needed is that nesting from yt.frontends import * is syntactically invalid.
| @@ -90,6 +90,8 @@ def load( | |||
| yt.utilities.exceptions.YTAmbiguousDataType | |||
| If the data format matches more than one class of similar specialization levels. | |||
| """ | |||
| from yt.frontends import _all # type: ignore [attr-defined] # noqa | |||
There was a problem hiding this comment.
Oh and one extra comment/question -- by nesting this import in the load and load_simulation calls, does that mean one could import yt and then directly instantiate a frontend without ever importing all the other frontends? That would actually be a pretty nice consequence beyond simply delaying the imports... might be worth a mention in the docs even. Could make a difference for batch processing scripts.
There was a problem hiding this comment.
indeed, I think with this patch you'd be able to short-circuit all frontends being loaded, but I think at this point it should be seen as an implementation detail rather than a feature (in particular because circumventing yt.load would require different code if your frontends isn't builtin, but lives in an extension)
neutrinoceros
dismissed stale reviews from chrishavlin and matthewturk
via
5ecad14
Co-authored-by: Chris Havlin <[email protected]>
|
@matthewturk this is ready ! |
PR Summary
Still pursuing my quest of making yt's CLI decently fast. This saves about 50 to 100ms overhead on startup time on my (relatively quick) laptop.