main.py

from fastapi import FastAPI, Query, HTTPException
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from pydantic import BaseModel
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import HTMLResponse
from fastapi.staticfiles import StaticFiles
import os
import torch

# -----------------------
# Set cache dirs (avoid Docker errors)
# -----------------------
# os.environ["HF_HOME"] = "/tmp"
# os.environ["TRANSFORMERS_CACHE"] = "/tmp"
# os.environ["TORCHINDUCTOR_CACHE_DIR"] = "/tmp/torch_inductor_cache"
# os.makedirs("/tmp/torch_inductor_cache", exist_ok=True)
os.environ["TORCH_HOME"] = "/tmp/torch_home"
os.environ["TORCHINDUCTOR_CACHE_DIR"] = "/tmp/torch_inductor_cache"
os.makedirs("/tmp/torch_home", exist_ok=True)
os.makedirs("/tmp/torch_inductor_cache", exist_ok=True)

# -----------------------
# Model Setup
# -----------------------
model_id = "LLM360/K2-Think"

print("Loading tokenizer and model...")
tokenizer = AutoTokenizer.from_pretrained(model_id, cache_dir="/tmp")

bnb_config = BitsAndBytesConfig(
    load_in_8bit=True  # 8-bit quantization
)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=bnb_config,
    device_map="auto",
    cache_dir="/tmp"
)
print("Model loaded!")

# -----------------------
# FastAPI Setup
# -----------------------
app = FastAPI(title="K2-Think QA API", description="Serving K2-Think Hugging Face model with FastAPI")

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

app.mount("/static", StaticFiles(directory="static"), name="static")

# -----------------------
# Request Schema
# -----------------------
class QueryRequest(BaseModel):
    question: str
    max_new_tokens: int = 50
    temperature: float = 0.7
    top_p: float = 0.9

# -----------------------
# Endpoints
# -----------------------
@app.get("/")
def home():
    return {"message": "Welcome to K2-Think QA API 🚀"}

@app.get("/health")
def health():
    return {"status": "ok"}

@app.get("/ask")
def ask(question: str = Query(...), max_new_tokens: int = Query(50)):
    try:
        inputs = tokenizer(question, return_tensors="pt")
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            do_sample=True,
            temperature=0.7,
            top_p=0.9,
            pad_token_id=tokenizer.eos_token_id,
            return_dict_in_generate=True
        )
        answer = tokenizer.decode(outputs.sequences[0], skip_special_tokens=True)
        return {"question": question, "answer": answer}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

@app.post("/predict")
def predict(request: QueryRequest):
    try:
        inputs = tokenizer(request.question, return_tensors="pt")
        outputs = model.generate(
            **inputs,
            max_new_tokens=request.max_new_tokens,
            do_sample=True,
            temperature=request.temperature,
            top_p=request.top_p,
            pad_token_id=tokenizer.eos_token_id,
            return_dict_in_generate=True
        )
        answer = tokenizer.decode(outputs.sequences[0], skip_special_tokens=True)
        return {"question": request.question, "answer": answer}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))