Repository Public
1"""Faux moteur pour les TESTS et le développement du frontend uniquement.23Il reproduit le protocole du serveur mlx_lm (mêmes champs, mêmes événements SSE, mêmes erreurs) sans modèle :4le texte produit est un écho déterministe. Il n'est jamais déployé ni importé par le code de production.56Comportements pilotables par le contenu du dernier message utilisateur :7 « [slow] » → 150 ms entre les tokens « [tool] » → émet un appel d'outil (si tools fournis)8 « [json] » → émet un objet JSON valide « [badjson] » → émet du texte non JSON9 « [think] » → émet d'abord du raisonnement « [boom] » → coupe la connexion en cours de flux10"""1112from __future__ import annotations1314import asyncio15import json16import time17import uuid18from typing import Any1920from fastapi import FastAPI, Request21from fastapi.responses import JSONResponse, StreamingResponse2223app = FastAPI()24STATE: dict[str, Any] = {"active": 0, "max_active": 0, "requests": 0, "cancelled": 0, "bodies": []}252627@app.get("/health")28async def health() -> dict[str, str]:29 return {"status": "ok"}303132@app.get("/internal/stats")33async def stats() -> dict[str, Any]:34 return {35 "ready": True,36 "healthy": True,37 "requests": STATE["requests"],38 "model": "fake-engine",39 "pid": 0,40 "uptime_s": 1,41 "warmup_s": 0.0,42 "mlx": {"active_bytes": 0, "peak_bytes": 0, "cache_bytes": 0},43 "prompt_cache": {"entries": 0, "bytes": 0},44 "decode_concurrency": 8,45 "prompt_concurrency": 4,46 "versions": {"mlx": "fake", "mlx_lm": "fake"},47 }484950def _plan(body: dict[str, Any]) -> tuple[list[dict[str, Any]], float, str]:51 """Retourne (deltas, délai, finish_reason)."""52 if "prompt" in body:53 words = [w + " " for w in f"echo: {body['prompt']}".split(" ")]54 return [{"text": w} for w in words], 0.0, "stop"55 last = next((m for m in reversed(body["messages"]) if m["role"] == "user"), {"content": ""})56 text = str(last.get("content") or "")57 delay = 0.15 if "[slow]" in text else 0.058 deltas: list[dict[str, Any]] = []59 finish = "stop"60 if "[think]" in text:61 deltas += [{"reasoning": "je "}, {"reasoning": "réfléchis"}]62 if "[tool]" in text and body.get("tools"):63 name = body["tools"][0]["function"]["name"]64 deltas.append(65 {66 "tool_calls": [67 {68 "function": {"name": name, "arguments": json.dumps({"city": "Québec"})},69 "type": "function",70 "id": str(uuid.uuid4()),71 "index": 0,72 }73 ]74 }75 )76 finish = "tool_calls"77 elif "[badjson]" in text and "rejetée" not in text:78 deltas += [{"content": "voici "}, {"content": "du texte"}]79 elif "[json]" in text or "rejetée" in text:80 deltas += [{"content": '{"answer": '}, {"content": "42}"}]81 else:82 deltas += [{"content": w + " "} for w in f"echo: {text}".split(" ")]83 limit = int(body.get("max_tokens", 512))84 if len(deltas) > limit:85 deltas, finish = deltas[:limit], "length"86 return deltas, delay, finish878889@app.post("/v1/chat/completions")90@app.post("/v1/completions")91async def completions(request: Request) -> Any:92 body = await request.json()93 STATE["requests"] += 194 STATE["bodies"].append(body)95 del STATE["bodies"][:-50]96 if body.get("model") != "default_model":97 return JSONResponse({"error": "model must be 'default_model'"}, status_code=400)98 is_chat = "messages" in body99 deltas, delay, finish = _plan(body)100 boom = is_chat and "[boom]" in json.dumps(body["messages"])101 cid = f"chatcmpl-{uuid.uuid4()}"102 prompt_tokens = len(json.dumps(body.get("messages") or body.get("prompt"))) // 4103104 def evt(choices: list[dict[str, Any]], **extra: Any) -> bytes:105 return (106 b"data: "107 + json.dumps(108 {109 "id": cid,110 "object": "chat.completion.chunk",111 "model": "default_model",112 "created": int(time.time()),113 "choices": choices,114 **extra,115 }116 ).encode()117 + b"\n\n"118 )119120 async def gen() -> Any:121 STATE["active"] += 1122 STATE["max_active"] = max(STATE["max_active"], STATE["active"])123 sent = 0124 try:125 yield b": keepalive 1/1\n\n"126 for d in deltas:127 if delay:128 await asyncio.sleep(delay)129 if boom and sent == 1:130 raise RuntimeError("boom")131 if is_chat:132 yield evt([{"index": 0, "finish_reason": None, "delta": {"role": "assistant", **d}}])133 else:134 yield evt([{"index": 0, "finish_reason": None, "text": d["text"]}])135 sent += 1136 if is_chat:137 yield evt([{"index": 0, "finish_reason": finish, "delta": {"role": "assistant"}}])138 else:139 yield evt([{"index": 0, "finish_reason": finish, "text": ""}])140 yield evt(141 [],142 usage={143 "prompt_tokens": prompt_tokens,144 "completion_tokens": sent,145 "total_tokens": prompt_tokens + sent,146 "prompt_tokens_details": {"cached_tokens": 2},147 },148 )149 yield b"data: [DONE]\n\n"150 except asyncio.CancelledError:151 STATE["cancelled"] += 1152 raise153 finally:154 STATE["active"] -= 1155156 return StreamingResponse(gen(), media_type="text/event-stream")157