app.py

import os
import time
import torch
import numpy as np
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download 
from fastapi import FastAPI, Request
from my_model import StoryPointIncrementModel
from transformers import AutoTokenizer

# Define the temporary cache directory for all Hugging Face operations
CACHE_DIR = "/tmp/hf"
MAX_RETRIES = 3 

# ----------------------------
# Hugging Face writable paths & Timeout Fix
# ----------------------------
os.environ["HF_HOME"] = CACHE_DIR
os.environ["HF_HUB_CACHE"] = CACHE_DIR
os.environ["TRANSFORMERS_CACHE"] = CACHE_DIR
os.environ["HF_DATASETS_CACHE"] = CACHE_DIR
os.makedirs(CACHE_DIR, exist_ok=True)

# FIX: Set a high download timeout (300 seconds)
os.environ["HF_DOWNLOAD_TIMEOUT"] = "300" 
# FIX: Enable LFS transfer protocols to potentially bypass network proxy issues
os.environ["HF_HUB_ENABLE_GATING"] = "1" # <-- NEW ATTEMPT TO FIX DOWNLOAD STALL

# CRITICAL FIX: Base model with hidden dimension of 384.
MODEL_BASE_NAME = "sentence-transformers/all-MiniLM-L6-v2" 

# --- AUTHENTICATION FIX START ---
HF_AUTH_TOKEN = os.environ.get("HF_TOKEN")
# --- AUTHENTICATION FIX END ---


# ----------------------------
# Load safetensors model and tokenizer
# ----------------------------
model_path = None
try:
    if HF_AUTH_TOKEN:
        print(f"DEBUG: HF_TOKEN successfully retrieved. Starts with: {HF_AUTH_TOKEN[:5]}...")
    else:
        print("DEBUG: HF_TOKEN is NOT set in environment variables.")

    # 1. Download the checkpoint file with retry logic
    for attempt in range(MAX_RETRIES):
        try:
            print(f"Attempting to download model (Attempt {attempt + 1}/{MAX_RETRIES})...")
            
            # *** CRITICAL FIX: TARGET THE NEW, SMALLER QUANTIZED FILE ***
            model_path = hf_hub_download(
                repo_id="AgileGenAI/JIRA-story-point-increment-predictor",
                filename="model_fp16.safetensors", # <-- TARGETING 45.5 MB FILE
                cache_dir=CACHE_DIR,
                token=HF_AUTH_TOKEN,
                force_download=True 
            )
            print("Model download succeeded.")
            break 
        except Exception as download_error:
            if attempt < MAX_RETRIES - 1:
                print(f"Download attempt {attempt + 1} failed: {download_error}. Retrying in 5 seconds...")
                time.sleep(5)
            else:
                raise download_error 

    if model_path is None:
        raise Exception("Failed to download model after all retries.")

    # 2. Load the state dictionary
    state_dict = load_file(model_path)
    
    # 3. Initialize the model
    model = StoryPointIncrementModel(model_name=MODEL_BASE_NAME, cache_dir=CACHE_DIR)
    
    print("Model initialized from quantized file.")

    # --- Critical Fix: Simplified Key Mapping to find the missing regressor ---
    new_state_dict = {}
    
    for k, v in state_dict.items():
        # NOTE: Since the file is pre-quantized, we don't need the messy in-code conversion logic.
        
        if k.startswith('bert.'): 
            new_state_dict[f'encoder.{k}'] = v
            
        elif 'embeddings.' in k or 'encoder.' in k or 'pooler.' in k:
            new_state_dict[f'encoder.{k}'] = v

        elif k.startswith('regressor.'):
            new_state_dict[k] = v
        
        elif k.startswith('classifier.') or k.startswith('linear.') or k.startswith('output.'): 
            new_state_dict[f'regressor.{k.split(".", 1)[1]}'] = v
        
        elif k == 'weight' or k == 'bias':
            new_state_dict[f'regressor.{k}'] = v

        else:
            new_state_dict[k] = v


    # Load the state dictionary. Setting strict=False is critical.
    model.load_state_dict(new_state_dict, strict=False)
    model.eval()
    
    # 4. Load the tokenizer (tokenizer is not affected by precision)
    tokenizer = AutoTokenizer.from_pretrained(MODEL_BASE_NAME, cache_dir=CACHE_DIR)

except Exception as e:
    print(f"An error occurred during model loading: {e}")
    model = None
    tokenizer = None


# ----------------------------
# FastAPI app
# ----------------------------
app = FastAPI()

@app.get("/")
def read_root():
    """Returns a simple status message for the root path."""
    return {
        "status": "Story Point Prediction API is running", 
        "message": "Use the /predict endpoint with a POST request and JSON body for predictions."
    }

@app.post("/predict")
async def predict(request: Request):
    if model is None or tokenizer is None:
        return {"error": "Model initialization failed. Cannot predict. Check startup logs for errors."}

    data = await request.json()
    description = data.get("description", "")
    summary = data.get("summary", "")

    # Combine description and summary into a single input text
    text_input = f"{summary} [SEP] {description}"

    # Tokenize the input text
    encoded_input = tokenizer(
        text_input, 
        return_tensors='pt', 
        padding='max_length', 
        truncation=True, 
        max_length=512
    )
    
    input_ids = encoded_input['input_ids']
    attention_mask = encoded_input['attention_mask']

    # Predict increment
    with torch.no_grad():
        # Pass input_ids and attention_mask to the model
        output = model(input_ids=input_ids, attention_mask=attention_mask)
        
        # The output is a tensor of shape [1, 1], extract and round the value
        story_point_increment = int(round(output.item()))

    return {"story_point_increment": story_point_increment}