app.py

# training_space/app.py (FastAPI Backend)
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
import subprocess
import os
import uuid
from huggingface_hub import HfApi, HfFolder
from fastapi.middleware.cors import CORSMiddleware
import logging

app = FastAPI()

# Configure Logging
logging.basicConfig(
    filename='training.log',
    filemode='a',
    format='%(asctime)s - %(levelname)s - %(message)s',
    level=logging.INFO
)

# CORS Configuration
origins = [
    "https://Vishwas1-LLMBuilderPro.hf.space",  # Replace with your Gradio frontend Space URL
    "http://localhost",  # For local testing
    "https://web.postman.co",
]

app.add_middleware(
    CORSMiddleware,
    allow_origins=origins,
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

# Define the expected payload structure
class TrainingRequest(BaseModel):
    task: str  # 'generation' or 'classification'
    model_params: dict
    model_name: str
    dataset_name: str  # The name of the existing Hugging Face dataset


# Root Endpoint
@app.get("/")
def read_root():
    return {
        "message": "Welcome to the Training Space API!",
        "instructions": "To train a model, send a POST request to /train with the required parameters."
    }

# Train Endpoint
@app.post("/train")
def train_model(request: TrainingRequest):
    try:
        logging.info(f"Received training request for model: {request.model_name}, Task: {request.task}")
        
        # Create a unique directory for this training session
        session_id = str(uuid.uuid4())
        session_dir = f"./training_sessions/{session_id}"
        os.makedirs(session_dir, exist_ok=True)
        
        # No need to save dataset content; use dataset_name directly
        dataset_name = request.dataset_name
        
        # Define the absolute path to train_model.py
        TRAIN_MODEL_PATH = os.path.join(os.path.dirname(__file__), "train_model.py")
        
        # Prepare the command to run the training script with dataset_name
        cmd = [
            "python", TRAIN_MODEL_PATH,
            "--task", request.task,
            "--model_name", request.model_name,
            "--dataset_name", dataset_name,  # Pass dataset_name instead of dataset file path
            "--num_layers", str(request.model_params.get('num_layers', 12)),
            "--attention_heads", str(request.model_params.get('attention_heads', 1)),
            "--hidden_size", str(request.model_params.get('hidden_size', 64)),
            "--vocab_size", str(request.model_params.get('vocab_size', 30000)),
            "--sequence_length", str(request.model_params.get('sequence_length', 512))
        ]
        
        # Start the training process as a background task in the root directory
        subprocess.Popen(cmd, cwd=os.path.dirname(__file__))
        
        logging.info(f"Training started for model: {request.model_name}, Session ID: {session_id}")
        
        return {"status": "Training started", "session_id": session_id}
    
    except Exception as e:
        logging.error(f"Error during training request: {str(e)}")
        raise HTTPException(status_code=500, detail=str(e))

    
    except Exception as e:
        logging.error(f"Error during training request: {str(e)}")
        raise HTTPException(status_code=500, detail=str(e))

# Optional: Status Endpoint
@app.get("/status/{session_id}")
def get_status(session_id: str):
    session_dir = f"./training_sessions/{session_id}"
    log_file = os.path.join(session_dir, "training.log")
    if not os.path.exists(log_file):
        raise HTTPException(status_code=404, detail="Session ID not found.")
    
    with open(log_file, "r", encoding="utf-8") as f:
        logs = f.read()
    
    return {"session_id": session_id, "logs": logs}