preserve training param

[kevin-mindverse]

Enhancement: Update retrain method to handle training parameters like start_process

Description

Currently, the retrain method lacks the ability to update training parameters in the same way as the start_process method. This enhancement adds parameter handling capabilities to the retrain method and modifies the TrainingParamsManager.update_training_params method to support resetting parameters when retraining.

Changes Made

• Enhanced the retrain endpoint to accept the same training parameters as start_process:
  • learning_rate
  • number_of_epochs
  • concurrency_threads
  • data_synthesis_mode
  • use_cuda
  • is_cot
• Modified the TrainingParamsManager.update_training_params method to accept a use_previous_params flag
• Set use_previous_params=False in the retrain method to ensure parameters are reset when retraining
• Updated endpoint documentation to reflect new parameters
• Added proper response data including all training parameters

Technical Details

• The retrain method now processes all training parameters from the request JSON
• Default values are applied for missing parameters
• When calling update_training_params, we pass use_previous_params=False to reset parameters
• The API response now includes all training parameters in the data section

Testing

1. Call the /api/trainprocess/retrain endpoint with a JSON payload containing:

   {
     "model_name": "Qwen2.5-0.5B-Instruct",
     "learning_rate": 2e-4,
     "number_of_epochs": 5,
     "concurrency_threads": 4,
     "data_synthesis_mode": "medium",
     "use_cuda": true,
     "is_cot": true
   }