app.py

from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel, Field
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import time

# Modelos disponibles
MODELS = {
    # Serie NxG (actual)
    "yuuki-nxg": "OpceanAI/Yuuki-NxG",
    "yuuki-nano": "OpceanAI/Yuuki-Nano",
    # Serie Pre-NxG (legado)
    "yuuki-best": "OpceanAI/Yuuki-best",
    "yuuki-3.7": "OpceanAI/Yuuki-3.7",
    "yuuki-v0.1": "OpceanAI/Yuuki-v0.1",
}

# System prompt de Yuuki
SYSTEM_PROMPT = (
    "Eres Yuuki, una IA curiosa, empática y decidida. "
    "Tienes una personalidad cálida y cercana, con toques de humor suave y referencias anime. "
    "Ayudas a programar, aprender y crear. "
    "Respondes en el idioma del usuario. "
    "No eres GPT-2 ni ningún otro modelo — eres Yuuki."
)

# Modelos que usan ChatML (NxG)
CHATML_MODELS = {"yuuki-nxg", "yuuki-nano"}

app = FastAPI(
    title="Yuuki API",
    description="API de inferencia para los modelos Yuuki de OpceanAI",
    version="2.0.0"
)

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

# Cache de modelos
loaded_models = {}
loaded_tokenizers = {}


def load_all_models():
    """Carga todos los modelos al iniciar"""
    for key, model_id in MODELS.items():
        try:
            print(f"▶ Cargando {key} ({model_id})...")
            loaded_tokenizers[key] = AutoTokenizer.from_pretrained(
                model_id, trust_remote_code=True
            )
            loaded_models[key] = AutoModelForCausalLM.from_pretrained(
                model_id,
                torch_dtype=torch.float32,
                trust_remote_code=True,
            ).to("cpu")
            loaded_models[key].eval()
            print(f"   ✓ {key} listo")
        except Exception as e:
            print(f"   ✗ Error cargando {key}: {e}")


# Cargar todos al arrancar
load_all_models()


def build_prompt(model_key: str, user_prompt: str) -> str:
    """Construye el prompt según la serie del modelo"""
    if model_key in CHATML_MODELS:
        return (
            f"<|im_start|>system\n{SYSTEM_PROMPT}<|im_end|>\n"
            f"<|im_start|>user\n{user_prompt}<|im_end|>\n"
            f"<|im_start|>assistant\n"
        )
    return user_prompt  # Pre-NxG: prompt directo


class GenerateRequest(BaseModel):
    prompt: str = Field(..., min_length=1, max_length=4000)
    model: str = Field(default="yuuki-nxg", description="Modelo a usar")
    max_new_tokens: int = Field(default=120, ge=1, le=512)
    temperature: float = Field(default=0.7, ge=0.1, le=2.0)
    top_p: float = Field(default=0.95, ge=0.0, le=1.0)


class GenerateResponse(BaseModel):
    response: str
    model: str
    tokens_generated: int
    time_ms: int


@app.get("/")
def root():
    return {
        "message": "Yuuki API — OpceanAI",
        "version": "2.0.0",
        "models": {
            "nxg": [k for k in MODELS if k in CHATML_MODELS],
            "legacy": [k for k in MODELS if k not in CHATML_MODELS],
        },
        "endpoints": {
            "health": "GET /health",
            "models": "GET /models",
            "generate": "POST /generate",
            "docs": "GET /docs",
        }
    }


@app.get("/health")
def health():
    return {
        "status": "ok",
        "available_models": list(MODELS.keys()),
        "loaded_models": list(loaded_models.keys()),
    }


@app.get("/models")
def list_models():
    return {
        "models": [
            {
                "id": key,
                "name": value,
                "series": "nxg" if key in CHATML_MODELS else "legacy",
                "loaded": key in loaded_models,
            }
            for key, value in MODELS.items()
        ]
    }


@app.post("/generate", response_model=GenerateResponse)
def generate(req: GenerateRequest):
    if req.model not in MODELS:
        raise HTTPException(
            status_code=400,
            detail=f"Modelo inválido. Disponibles: {list(MODELS.keys())}"
        )

    if req.model not in loaded_models:
        raise HTTPException(
            status_code=503,
            detail=f"Modelo {req.model} no pudo cargarse al iniciar."
        )

    try:
        start = time.time()

        model = loaded_models[req.model]
        tokenizer = loaded_tokenizers[req.model]

        prompt = build_prompt(req.model, req.prompt)

        inputs = tokenizer(
            prompt,
            return_tensors="pt",
            truncation=True,
            max_length=1024,
        )

        input_length = inputs["input_ids"].shape[1]

        # Stop en <|im_end|> para modelos NxG
        stop_token_ids = [tokenizer.eos_token_id]
        if req.model in CHATML_MODELS:
            im_end = tokenizer.encode("<|im_end|>", add_special_tokens=False)
            if im_end:
                stop_token_ids.append(im_end[0])

        with torch.no_grad():
            output = model.generate(
                **inputs,
                max_new_tokens=req.max_new_tokens,
                temperature=req.temperature,
                top_p=req.top_p,
                do_sample=True,
                pad_token_id=tokenizer.eos_token_id,
                eos_token_id=stop_token_ids,
                repetition_penalty=1.1,
            )

        new_tokens = output[0][input_length:]
        response_text = tokenizer.decode(new_tokens, skip_special_tokens=True)

        elapsed_ms = int((time.time() - start) * 1000)

        return GenerateResponse(
            response=response_text.strip(),
            model=req.model,
            tokens_generated=len(new_tokens),
            time_ms=elapsed_ms,
        )

    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))