fix: ignore "train" part of schema errors during predict runs #1496
Conversation
dkhokhlov
commented
Jan 24, 2024
•
dkhokhlov
commented
- ignore "train" part of schema errors during predict runs for backward compatibility with existing "bad" models already in use
- expected: cog shall not trigger exception/error when "train" is present in config but "train.py" is missing.
dkhokhlov
force-pushed
the
ignore_schema_exceptions_on_trainer_path
branch
2 times, most recently
from
8daf55d
to
6a65211
Compare
dkhokhlov
changed the title
There was a problem hiding this comment.
Nice work, @dkhokhlov. Thanks for taking this on.
Left a few non-blocking suggestions for renaming predictor_ref to training_ref. But other than that, I think this is solid.
| mode: str = "predict", | ||
| is_build: bool = False |
There was a problem hiding this comment.
I like the decision to define this behavior in a new argument to create_app. We originally discussed defining a new mode value, but I see the value of using something that doesn't map to a command-line argument, and therefore risk misuse (e.g. What's the expected behavior of a pod running a model with the build mode?)
There was a problem hiding this comment.
I'm not wild about this particular spelling, but I can't think of anything better. Since it's effectively a private API, it's probably fine. Or else, we can change it later without breaking anyone.
…uction for backward compatibility Signed-off-by: Dmitri Khokhlov <dkhokhlov@gmail.com>
dkhokhlov
force-pushed
the
ignore_schema_exceptions_on_trainer_path
branch
from
6a65211
to
d60370e
Compare
