app.py

from fastapi import FastAPI, Query
from pydantic import BaseModel
import cloudscraper
from bs4 import BeautifulSoup
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import re
import os
os.environ["TRANSFORMERS_CACHE"] = "/tmp/.cache"
os.environ["HF_HOME"] = "/tmp/.cache"
os.environ["HF_DATASETS_CACHE"] = "/tmp/.cache"

app = FastAPI()

class ThreadResponse(BaseModel):
    question: str
    replies: list[str]

def clean_text(text: str) -> str:
    text = text.strip()
    text = re.sub(r"\b\d+\s*likes?,?\s*\d*\s*replies?$", "", text, flags=re.IGNORECASE).strip()
    return text

@app.get("/scrape", response_model=ThreadResponse)
def scrape(url: str = Query(...)):
    scraper = cloudscraper.create_scraper()
    response = scraper.get(url)

    if response.status_code == 200:
        soup = BeautifulSoup(response.content, 'html.parser')
        comment_containers = soup.find_all('div', class_='post__content')

        if comment_containers:
            question = clean_text(comment_containers[0].get_text(strip=True, separator="\n"))
            replies = [clean_text(comment.get_text(strip=True, separator="\n")) for comment in comment_containers[1:]]
            return ThreadResponse(question=question, replies=replies)
    return ThreadResponse(question="", replies=[])


MODEL_NAME = "sarvamai/sarvam-m"

# Load tokenizer and model once at startup, with device auto-mapping
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, torch_dtype="auto", device_map="auto")
model.eval()

class PromptRequest(BaseModel):
    prompt: str

@app.post("/generate")
async def generate_text(request: PromptRequest):
    # Prepare chat-style input with thinking mode enabled
    messages = [{"role": "user", "content": request.prompt}]
    text = tokenizer.apply_chat_template(messages, tokenize=False, enable_thinking=True)

    inputs = tokenizer([text], return_tensors="pt").to(model.device)
    with torch.no_grad():
        generated_ids = model.generate(**inputs, max_new_tokens=512, temperature=0.5)
    output_ids = generated_ids[:, inputs.input_ids.shape[-1]:].tolist()[0]
    output_text = tokenizer.decode(output_ids)

    # Extract reasoning and content parts if thinking tags are present
    if "</think>" in output_text:
        reasoning_content = output_text.split("</think>")[0].strip()
        content = output_text.split("</think>")[1].strip().rstrip("</s>")
    else:
        reasoning_content = ""
        content = output_text.strip().rstrip("</s>")

    return {
        "reasoning_content": reasoning_content,
        "generated_text": content
    }