api.py

"""
API FastAPI pour DeepSeek-Coder avec llama_cpp
Utilise des modèles publics accessibles
"""
import os
import time
import asyncio
from typing import Optional, List
from contextlib import asynccontextmanager

from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel, Field

# Import llama_cpp
try:
    from llama_cpp import Llama
except ImportError:
    Llama = None

try:
    from huggingface_hub import hf_hub_download
    HF_AVAILABLE = True
except ImportError:
    HF_AVAILABLE = False

# ========== CONFIGURATION ==========
# Dépôts PUBLICs et accessibles
MODEL_CONFIGS = [
    {
        "repo": "TheBloke/DeepSeek-Coder-1.3B-Instruct-GGUF",
        "files": [
            "deepseek-coder-1.3b-instruct.Q4_K_M.gguf",
            "deepseek-coder-1.3b-instruct.Q4_0.gguf",
            "deepseek-coder-1.3b-instruct.Q2_K.gguf"
        ]
    },
    {
        "repo": "mradermacher/DeepSeek-Coder-1.3B-Instruct-GGUF",
        "files": [
            "DeepSeek-Coder-1.3B-Instruct.Q4_K_M.gguf",
            "DeepSeek-Coder-1.3B-Instruct.Q2_K.gguf"
        ]
    }
]

# Modèle de secours plus petit
FALLBACK_MODELS = [
    {
        "repo": "TheBloke/CodeLlama-7B-Instruct-GGUF",
        "files": ["codellama-7b-instruct.Q2_K.gguf"]  # ~2.7GB
    },
    {
        "repo": "TheBloke/tinycoder-1.1B-GGUF",
        "files": ["tinycoder-1.1b.Q2_K.gguf"]  # ~500MB
    }
]

MODEL_DIR = "./models"
os.makedirs(MODEL_DIR, exist_ok=True)

# ========== MODÈLES DE DONNÉES ==========
class GenerateRequest(BaseModel):
    prompt: str = Field(..., min_length=1, max_length=2000)
    temperature: float = Field(0.2, ge=0.1, le=1.0)
    max_tokens: int = Field(256, ge=1, le=1024)
    top_p: float = Field(0.95, ge=0.1, le=1.0)

class ChatMessage(BaseModel):
    role: str = Field(..., pattern="^(user|assistant|system)$")
    content: str

class ChatRequest(BaseModel):
    messages: List[ChatMessage]
    temperature: float = Field(0.2, ge=0.1, le=1.0)
    max_tokens: int = Field(256, ge=1, le=1024)
# ========== OPENAI / OPENROUTER SCHEMAS ==========

class OpenAIMessage(BaseModel):
    role: str
    content: str

class ChatCompletionRequest(BaseModel):
    model: str
    messages: List[OpenAIMessage]
    temperature: Optional[float] = 0.2
    max_tokens: Optional[int] = 256
    top_p: Optional[float] = 0.95
    stream: Optional[bool] = False

class CompletionRequest(BaseModel):
    model: str
    prompt: str
    temperature: Optional[float] = 0.2
    max_tokens: Optional[int] = 256
    top_p: Optional[float] = 0.95

# ========== GESTION DU MODÈLE ==========
class ModelManager:
    def __init__(self):
        self.llm = None
        self.model_path = None
        self.loading = False
        self.model_loaded = False
        
    def download_model(self, repo_id: str, filename: str) -> str:
        """Télécharger un modèle depuis Hugging Face"""
        print(f"📥 Téléchargement: {filename} depuis {repo_id}")
        
        try:
            model_path = hf_hub_download(
                repo_id=repo_id,
                filename=filename,
                local_dir=MODEL_DIR,
                local_dir_use_symlinks=False,
                resume_download=True,
                token=None  # Pas de token nécessaire pour les repos publics
            )
            print(f"✅ Téléchargé: {model_path}")
            return model_path
        except Exception as e:
            print(f"❌ Erreur: {str(e)[:200]}")
            raise
    
    def find_or_download_model(self):
        """Trouver ou télécharger un modèle accessible"""
        if not HF_AVAILABLE:
            raise Exception("huggingface-hub non disponible")
        
        # Vérifier les modèles existants
        for root, dirs, files in os.walk(MODEL_DIR):
            for file in files:
                if file.endswith('.gguf'):
                    path = os.path.join(root, file)
                    print(f"✅ Modèle existant trouvé: {path}")
                    return path
        
        # Essayer les modèles principaux
        print("🔍 Recherche d'un modèle accessible...")
        
        for config in MODEL_CONFIGS:
            repo = config["repo"]
            for filename in config["files"]:
                try:
                    return self.download_model(repo, filename)
                except:
                    continue
        
        # Essayer les modèles de secours
        print("🔄 Essai des modèles de secours...")
        for config in FALLBACK_MODELS:
            repo = config["repo"]
            for filename in config["files"]:
                try:
                    return self.download_model(repo, filename)
                except:
                    continue
        
        # Si tout échoue, créer un modèle factice pour tester
        print("⚠️ Création d'un modèle factice pour test...")
        dummy_path = os.path.join(MODEL_DIR, "dummy.gguf")
        with open(dummy_path, 'w') as f:
            f.write("DUMMY MODEL FOR TESTING")
        
        # Dans un environnement réel, vous voudriez télécharger un vrai petit modèle
        # Exemple: "microsoft/phi-2" ou un petit modèle local
        return dummy_path
    
    def load_model(self):
        """Charger le modèle avec llama_cpp"""
        if self.llm is not None:
            return self.llm
        
        if Llama is None:
            raise Exception("llama_cpp n'est pas installé")
            
        print("🔧 Chargement du modèle...")
        self.loading = True
        
        try:
            # Trouver ou télécharger le modèle
            self.model_path = self.find_or_download_model()
            
            # Vérifier si c'est un modèle factice
            if "dummy" in self.model_path:
                print("⚠️ Utilisation du modèle factice - l'API fonctionnera en mode test")
                self.llm = DummyLLM()
                self.model_loaded = True
                self.loading = False
                return self.llm
            
            # Configuration optimisée
            n_gpu_layers = 0  # Pas de GPU sur Hugging Face gratuit
            n_threads = 2     # Conservateur
            n_ctx = 2048      # Limité
            
            print(f"🔄 Chargement: {os.path.basename(self.model_path)}")
            
            # Charger le vrai modèle
            self.llm = Llama(
                model_path=self.model_path,
                n_ctx=n_ctx,
                n_threads=n_threads,
                n_gpu_layers=n_gpu_layers,
                verbose=False
            )
            
            print("✅ Modèle chargé!")
            self.model_loaded = True
            self.loading = False
            return self.llm
            
        except Exception as e:
            print(f"❌ Erreur: {e}")
            print("🔄 Utilisation du mode fallback...")
            self.llm = DummyLLM()
            self.model_loaded = True
            self.loading = False
            return self.llm
    
    def generate(
        self,
        prompt: str,
        temperature: float = 0.2,
        max_tokens: int = 256,
        top_p: float = 0.95
    ):
        """Générer du texte avec llama_cpp"""
        if self.llm is None:
            self.load_model()

        if isinstance(self.llm, DummyLLM):
            return self.llm.generate(prompt, temperature, max_tokens, top_p)

        try:
            response = self.llm(
                prompt,
                max_tokens=max_tokens,
                temperature=temperature,
                top_p=top_p,
                echo=False
            )

            return response["choices"][0]["text"]

        except Exception as e:
            return (
                "# Fallback response\n\n"
                f"Prompt: {prompt}\n\n"
                f"Error: {str(e)[:200]}"
            )
        
    def chat(
        self,
        messages: List[dict],
        temperature: float = 0.2,
        max_tokens: int = 256
    ):
        if self.llm is None:
            self.load_model()

        if isinstance(self.llm, DummyLLM):
            return self.llm.chat(messages, temperature, max_tokens)

        # ---- Prompt Instruct DeepSeek ----
        system_prompt = ""
        user_prompt = ""

        for msg in messages:
            if msg["role"] == "system":
                system_prompt += msg["content"] + "\n"
            elif msg["role"] == "user":
                user_prompt += msg["content"] + "\n"

        prompt = f"""### Instruction:
        {system_prompt}{user_prompt}

        ### Response:
        """

        response = self.llm(
            prompt,
            max_tokens=max_tokens,
            temperature=temperature,
            top_p=0.95,
            echo=False
        )

        return response["choices"][0]["text"].strip()





# ========== MODÈLE FACTICE POUR TEST ==========
class DummyLLM:
    """Modèle factice pour tester l'API quand le vrai modèle échoue"""
    
    def generate(self, prompt: str, temperature: float = 0.2, max_tokens: int = 256, top_p: float = 0.95):
        """Générer une réponse factice"""
        # Simuler un délai de traitement
        time.sleep(0.1)
        
        # Retourner une réponse basée sur le prompt
        if "python" in prompt.lower():
            return f"""# Code Python généré (mode test)
# Prompt: {prompt[:50]}...

def example_function():
    \"\"\"Exemple de fonction Python\"\"\"
    print("Hello from DeepSeek-Coder (Test Mode)")
    return 42

# Note: L'API fonctionne mais utilise un modèle factice.
# Le vrai modèle sera téléchargé automatiquement à la prochaine requête."""
        
        elif "javascript" in prompt.lower() or "js" in prompt.lower():
            return f"""// Code JavaScript généré (mode test)
// Prompt: {prompt[:50]}...

function exampleFunction() {{
    console.log("Hello from DeepSeek-Coder (Test Mode)");
    return 42;
}}

// Note: Mode test - le vrai modèle se télécharge en arrière-plan."""
        
        else:
            return f"""# Réponse générée (mode test)
Prompt: {prompt}

Voici un exemple de code:
```python
def process_input(text):
    \"\"\"Traiter l'entrée utilisateur\"\"\"
    return f"Processed: {{text}}"

# L'API est opérationnelle en mode test.
# Le modèle DeepSeek-Coder se télécharge en arrière-plan."""

    def chat(self, messages: List[dict], temperature: float = 0.2, max_tokens: int = 256):
        """Chat factice"""
        last_message = messages[-1]["content"] if messages else "Hello"
        
        responses = [
            f"Bonjour! Je suis DeepSeek-Coder en mode test. Vous avez dit: '{last_message[:50]}...'",
            f"Je peux vous aider avec du code. En mode test, voici un exemple:\n\n```python\nprint('Hello World')\n```",
            f"Le modèle réel est en cours de téléchargement. En attendant, voici une réponse de test."
        ]
        
        import random
        return random.choice(responses)

# ========== APPLICATION ==========
model_manager = ModelManager()

@asynccontextmanager
async def lifespan(app: FastAPI):
    """Cycle de vie"""
    print("🚀 Démarrage API...")
    
    async def load_async():
        try:
            model_manager.load_model()
        except Exception as e:
            print(f"⚠️ Note: {e}")
    
    asyncio.create_task(load_async())
    yield
    
    print("🧹 Arrêt...")

app = FastAPI(
    title="🚀 DeepSeek-Coder API",
    description="API avec fallback automatique",
    version="1.0.0",
    docs_url="/docs",
    redoc_url=None,
    lifespan=lifespan
)

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

# ========== ROUTES ==========
@app.get("/")
async def root():
    return {
        "message": "🚀 DeepSeek-Coder 1.3B API",
        "status": "ready" if model_manager.model_loaded else "loading",
        "mode": "dummy" if isinstance(model_manager.llm, DummyLLM) else "real",
        "endpoints": {
            "generate": "POST /generate",
            "chat": "POST /chat",
            "health": "GET /health",
            "test": "GET /test"
        }
    }

@app.get("/health")
async def health():
    return {
        "status": "healthy",
        "model_loaded": model_manager.model_loaded,
        "model_type": "dummy" if isinstance(model_manager.llm, DummyLLM) else "real",
        "timestamp": time.time()
    }

@app.post("/generate")
async def generate(request: GenerateRequest):
    """Générer du code"""
    try:
        response = model_manager.generate(
            prompt=request.prompt,
            temperature=request.temperature,
            max_tokens=request.max_tokens,
            top_p=request.top_p
        )
        
        return {
            "response": response,
            "model": "deepseek-coder-1.3b",
            "mode": "dummy" if isinstance(model_manager.llm, DummyLLM) else "real",
            "backend": "llama_cpp"
        }
        
    except Exception as e:
        # Fallback encore plus simple
        return {
            "response": f"# Fallback response\n\nPrompt: {request.prompt}\n\nError: {str(e)[:100]}",
            "model": "fallback",
            "error": "generation_failed"
        }

@app.post("/chat")
async def chat(request: ChatRequest):
    """Chat"""
    try:
        messages = [msg.dict() for msg in request.messages]
        response = model_manager.chat(
            messages=messages,
            temperature=request.temperature,
            max_tokens=request.max_tokens
        )
        
        return {
            "response": response,
            "model": "deepseek-coder-1.3b-instruct",
            "mode": "dummy" if isinstance(model_manager.llm, DummyLLM) else "real"
        }
        
    except Exception as e:
        return {
            "response": f"Chat error: {str(e)[:100]}",
            "model": "fallback"
        }

@app.get("/test")
async def test():
    """Endpoint de test"""
    return {
        "test": "success",
        "message": "API is running",
        "try_endpoints": [
            "POST /generate with JSON: {'prompt': 'def hello():', 'temperature': 0.2}",
            "POST /chat with JSON: {'messages': [{'role': 'user', 'content': 'Hello'}]}"
        ]
    }

@app.get("/download_status")
async def download_status():
    """Vérifier le statut du téléchargement"""
    if model_manager.model_path and os.path.exists(model_manager.model_path):
        size = os.path.getsize(model_manager.model_path)
        return {
            "downloaded": True,
            "size_mb": round(size / 1024 / 1024, 2),
            "path": model_manager.model_path
        }
    return {"downloaded": False, "message": "No model downloaded yet"}

@app.post("/v1/chat/completions")
async def openai_chat(request: ChatCompletionRequest):
    try:
        messages = [m.dict() for m in request.messages]

        response_text = model_manager.chat(
            messages=messages,
            temperature=request.temperature,
            max_tokens=request.max_tokens
        )

        return {
            "id": f"chatcmpl-{int(time.time())}",
            "object": "chat.completion",
            "created": int(time.time()),
            "model": request.model,
            "choices": [
                {
                    "index": 0,
                    "message": {
                        "role": "assistant",
                        "content": response_text
                    },
                    "finish_reason": "stop"
                }
            ],
            "usage": {
                "prompt_tokens": 0,
                "completion_tokens": 0,
                "total_tokens": 0
            }
        }

    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))
@app.post("/v1/completions")
async def openai_completion(request: CompletionRequest):
    try:
        response_text = model_manager.generate(
            prompt=request.prompt,
            temperature=request.temperature,
            max_tokens=request.max_tokens,
            top_p=request.top_p
        )

        return {
            "id": f"cmpl-{int(time.time())}",
            "object": "text_completion",
            "created": int(time.time()),
            "model": request.model,
            "choices": [
                {
                    "text": response_text,
                    "index": 0,
                    "finish_reason": "stop"
                }
            ]
        }

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

if __name__ == "__main__":
    import uvicorn
    port = int(os.getenv("PORT", 7860))
    print(f"🌐 Serveur démarré sur le port {port}")
    uvicorn.run(app, host="0.0.0.0", port=port)