Model Registry

[ChakshuGautam]

• Model Version Management (commit hash, semantic version) - should happen while training
• Provide model files (onnx, pt, bin) through a CDN
• Rollback to an older version
• Deployment by a version number
• Track costs during training

Clicking train button on Admi Panel

ML Pod:

• Modify the train API to support versioning #142

Admin panel :

• when train button is clicked ,it'll hit model registry API to get the:

  • Base Model Branch on HF - the base model which will be used to train the dataset with
  • task_type: classfication/NER etc
  • model_format: onnx/pytorch - safetensors
  • model_name (purpose for which model is getting trained ) like agri_classification in AKAI/KMAI {can be same as service_name}
  • epochs (number of epochs the model is getting trained for)
  • args : training arguements used to fine tune the model
  • quantization: None mostly unless specified)

• Admin Panel will hit dataset registry to get dataset id for the given model-botid

• Admin Panel will hit /train API with the following parameters:

{
    "model": Base Model Branch on HF (from model registry)
    "epochs":  (from model registry)
      "task_type":  (from model registry)
    "dataset":  (from dataset registry)
        "versioning": {
         "owner": botid   
        "environment": bot environment 
        “model_name ': (from model registry) 
    },
“args”: (from model registry) 
}

Dataset service:

• To create dataset for models with the following for each model-botid t least :

  • Base Model Branch on HF - the base model which will be used to train the dataset with
  • task_type: classfication/NER etc
  • model_format: onnx/pytorch - safetensors
  • model_name (purpose for which model is getting trained ) like agri_classification in AKAI/KMAI {can be same as service_name}
  • epochs (number of epochs the model is getting trained for)
  • args : training arguements used to fine tune the model
  • quantization: None mostly unless specified)

• to create dataset for datasets with :
  datasetid for each model for each bot

[ChakshuGautam]

@suresh12 to review the Doc

[Gautam-Rajeev]

Clicking train button on Admi Panel

ML Pod:

• Modify the train API to support versioning

Admin panel :

• when train button is clicked ,it'll hit model registry API to get the:

  • Base Model Branch on HF - the base model which will be used to train the dataset with
  • task_type: classfication/NER etc
  • model_format: onnx/pytorch - safetensors
  • model_name (purpose for which model is getting trained ) like agri_classification in AKAI/KMAI {can be same as service_name}
  • epochs (number of epochs the model is getting trained for)
  • args : training arguements used to fine tune the model
  • quantization: None mostly unless specified)

• Admin Panel will hit dataset registry to get dataset id for the given model-botid

• Admin Panel will hit /train API with the following parameters:

{
    "model": Base Model Branch on HF (from model registry)
    "epochs":  (from model registry)
      "task_type":  (from model registry)
    "dataset":  (from dataset registry)
        "versioning": {
         "owner": botid   
        "environment": bot environment 
        “model_name ': (from model registry) 
    },
“args”: (from model registry) 
}

Dataset service:

• To create dataset for models with the following for each model-botid t least :

  • Base Model Branch on HF - the base model which will be used to train the dataset with
  • task_type: classfication/NER etc
  • model_format: onnx/pytorch - safetensors
  • model_name (purpose for which model is getting trained ) like agri_classification in AKAI/KMAI {can be same as service_name}
  • epochs (number of epochs the model is getting trained for)
  • args : training arguements used to fine tune the model
  • quantization: None mostly unless specified)

• to create dataset for datasets with :
  datasetid for each model for each bot

[KDwevedi]

Scoping Model Registry from ML Flow us lift and use directly.

Desirable Features

• Model Metadata Storage
• Version Management + Finetuning
• Deployment
• Utilising CDNs for making BIN and other model files available
• Metrics for model use recorded
• PoC