Compter des tokens (tiktoken)

import tiktoken

enc = tiktoken.encoding_for_model("gpt-4o")
text = "Bonjour le monde, comment ça va ?"
tokens = enc.encode(text)
print(f"Tokens: {len(tokens)}")        # ~10 tokens
print(f"Decoded: {[enc.decode([t]) for t in tokens]}")

Servir Llama 3 8B avec vLLM

# Installation
pip install vllm

# Lancement serveur (OpenAI-compatible API)
vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
  --host 0.0.0.0 --port 8000 \
  --max-model-len 8192 \
  --tensor-parallel-size 1 \
  --gpu-memory-utilization 0.92

Appel client (compatible OpenAI SDK)

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")
resp = client.chat.completions.create(
    model="meta-llama/Meta-Llama-3-8B-Instruct",
    messages=[
        {"role": "system", "content": "Tu es un assistant utile."},
        {"role": "user", "content": "Résume Apache Spark en 2 phrases."}
    ],
    max_tokens=200,
    temperature=0.3,
    stream=True
)
for chunk in resp:
    print(chunk.choices[0].delta.content, end="")

Wrapper FastAPI avec rate limiting, auth, logging

from fastapi import FastAPI, Depends, HTTPException
from fastapi.security import APIKeyHeader
from pydantic import BaseModel
import httpx
from slowapi import Limiter
from slowapi.util import get_remote_address

app = FastAPI()
limiter = Limiter(key_func=get_remote_address)
api_key_header = APIKeyHeader(name="X-API-Key")

class ChatReq(BaseModel):
    messages: list[dict]
    max_tokens: int = 500
    temperature: float = 0.3

def auth(key: str = Depends(api_key_header)):
    if key not in VALID_KEYS:
        raise HTTPException(401, "Invalid API key")

@app.post("/chat")
@limiter.limit("30/minute")
async def chat(req: ChatReq, _=Depends(auth)):
    async with httpx.AsyncClient() as c:
        r = await c.post("http://vllm:8000/v1/chat/completions",
                         json=req.dict(), timeout=60)
    return r.json()