[AQUA] Added support for deploy stack model. #1223
Conversation
oracle-contributor-agreement
bot
added
the
OCA Verified
mrDzurb
changed the title
lu-ohai
requested review from
darenr,
mayoor,
mrDzurb,
VipulMascarenhas,
qiuosier and
ahosler
as code owners
lu-ohai
changed the title
ads/aqua/model/model.py
Outdated
| """ | ||
| fine_tune_weights = ( |
There was a problem hiding this comment.
looks like there are several place where we do same validation. Wouldn't it be better to have something like:
if isinstance(model_id, AquaMultiModelRef):
....
else:
....
| f"Aqua Model {custom_model.id} created with the service model {model_id}." | ||
| ) | ||
| custom_model = None | ||
| if fine_tune_weights: |
There was a problem hiding this comment.
Wouldn’t it be better, instead of handling everything in one place, to split the implementation into two separate methods, one for creating the MC record and another for creating the group?
There was a problem hiding this comment.
Because creating model or model group will share some configs, like compartment/project/all tags, that's why I put it in one method. I just created separate helper functions for each creation.
| if create_deployment_details.model_id: | ||
| if ( | ||
| create_deployment_details.model_id | ||
| or create_deployment_details.deployment_type == DEFAULT_DEPLOYMENT_TYPE |
There was a problem hiding this comment.
I think this is a bit dangerous, what if someone change the default deployment type?
There was a problem hiding this comment.
We can add a validation for it and only MODEL_STACK is supported at this moment.
| @@ -206,13 +214,28 @@ def create( | |||
|
|
|||
| # Create an AquaModelApp instance once to perform the deployment creation. | |||
| model_app = AquaModelApp() | |||
| if create_deployment_details.model_id: | |||
| if ( | |||
| create_deployment_details.model_id | |||
There was a problem hiding this comment.
I think this assumption is a bit wrong, no?
create_deployment_details.model_id can be provided only for a single model deployment. In case of MMD or Stacked deployment we should rely on create_deployment_details.models, no?
There was a problem hiding this comment.
I think we talked about this before, that is for single dployment, we adopt model_id and for stacked deployment, we adopt models with deployment_type. This assumption will filter out both single and stack deployment cases.
| ) | ||
|
|
||
| build_model_group_details = copy.deepcopy(self._spec) | ||
| build_model_group_details.pop(self.CONST_CUSTOM_METADATA_LIST) | ||
| build_model_group_details.pop(self.CONST_MEMBER_MODELS) | ||
| build_model_group_details.pop(self.CONST_CUSTOM_METADATA_LIST, None) |
There was a problem hiding this comment.
Add a comment here, why we are doing this?
| is not None | ||
| or service_model.freeform_tags.get(Tags.AQUA_FINE_TUNED_MODEL_TAG) | ||
| is not None | ||
| ): |
There was a problem hiding this comment.
NIT: if (
service_model.freeform_tags.get(Tags.BASE_MODEL_CUSTOM) is None and
service_model.freeform_tags.get(Tags.AQUA_FINE_TUNED_MODEL_TAG) is None
):
| for fine_tune_weight in fine_tune_weights | ||
| ] | ||
| ) | ||
| .create() |
There was a problem hiding this comment.
so is this only for 1 FT model + multiple LoRA modules?
There was a problem hiding this comment.
Only one base model + multiple LoRA modules
| @@ -147,9 +147,9 @@ def create( | |||
| freeform_tags: Optional[Dict] = None, | |||
| defined_tags: Optional[Dict] = None, | |||
| **kwargs, | |||
| ) -> DataScienceModel: | |||
| ) -> Union[DataScienceModel, DataScienceModelGroup]: | |||
There was a problem hiding this comment.
Myabe we can rename model_id to model now, it would make sense for me, but I guess will require some changes to the unit tests.
Added support for deploy stack model.
MODEL_STACKas single model deployment using model groupmodelsanddeployment_type=MODEL_STACKtocreateat the same time will trigger this type of deploymentMODELwill be added and the--served-model-nameinPARAMSwill be updated to the model id.Integration tests
Unit tests