feat: mlx-proxy 서버 + n8n 워크플로우 LLM/임베딩 URL 분리

mlx-vlm 기반 ollama 호환 프록시 서버 추가 (port 11435).
n8n GEN 노드 6개에 callLLM 래퍼 주입 (health check + ollama fallback).
임베딩/리랭커는 ollama(LOCAL_EMBED_URL)로 분리.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

mlx_proxy.py

@@ -0,0 +1,113 @@
+"""mlx-vlm proxy — ollama 호환 API for mlx-vlm inference (port 11435)
+
+ollama 호환 /api/generate 엔드포인트 제공. n8n에서 투명하게 사용 가능.
+Qwen3.5-27B-4bit를 mlx-vlm으로 서빙, thinking 자동 비활성화 (prefill 방식).
+"""
+
+import asyncio
+import logging
+import time
+from functools import partial
+
+from fastapi import FastAPI, Request
+from fastapi.responses import JSONResponse
+
+logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
+logger = logging.getLogger("mlx_proxy")
+
+MODEL_PATH = "mlx-community/Qwen3.5-27B-4bit"
+DISPLAY_NAME = "qwen3.5:27b"
+
+app = FastAPI()
+
+# Global model state — loaded once at startup
+model = None
+processor = None
+
+
+@app.on_event("startup")
+async def startup():
+    global model, processor
+    logger.info(f"Loading {MODEL_PATH} ...")
+    from mlx_vlm import load
+    model, processor = load(MODEL_PATH)
+    logger.info("Model ready")
+
+
+def _generate_sync(
+    prompt: str, system: str, max_tokens: int, temperature: float, is_json: bool
+) -> dict:
+    """Blocking generation — runs in thread pool executor."""
+    from mlx_vlm import generate as mlx_generate
+    from mlx_vlm.utils import apply_chat_template
+
+    messages = []
+
+    # System prompt — strip /no_think (thinking handled by prefill)
+    sys_text = (system or "").replace("/no_think", "").strip()
+    if is_json and sys_text and "JSON" not in sys_text:
+        sys_text += "\nJSON으로만 응답하세요."
+    elif is_json and not sys_text:
+        sys_text = "JSON으로만 응답하세요."
+    if sys_text:
+        messages.append({"role": "system", "content": sys_text})
+
+    messages.append({"role": "user", "content": prompt})
+    # Prefill: thinking 비활성화 — 모델이 <think> 단계 완료 상태에서 시작
+    messages.append({"role": "assistant", "content": "<think>\n</think>\n\n"})
+
+    formatted = apply_chat_template(processor, model.config, messages)
+
+    t0 = time.perf_counter()
+    output = mlx_generate(
+        model, processor, formatted, max_tokens=max_tokens, temp=temperature
+    )
+    elapsed = time.perf_counter() - t0
+
+    output = output.strip()
+
+    # Token count (approximate via tokenizer)
+    try:
+        n_tokens = len(processor.tokenizer.encode(output))
+    except Exception:
+        n_tokens = max(1, len(output) // 3)
+
+    return {"text": output, "n_tokens": n_tokens, "elapsed": elapsed}
+
+
+@app.post("/api/generate")
+async def api_generate(request: Request):
+    body = await request.json()
+
+    prompt = body.get("prompt", "")
+    system = body.get("system", "")
+    opts = body.get("options", {})
+    max_tokens = opts.get("num_predict", 2048)
+    temperature = opts.get("temperature", 0.7)
+    is_json = body.get("format") == "json"
+
+    if not prompt:
+        return JSONResponse({"error": "prompt required"}, status_code=400)
+
+    try:
+        loop = asyncio.get_event_loop()
+        result = await loop.run_in_executor(
+            None,
+            partial(_generate_sync, prompt, system, max_tokens, temperature, is_json),
+        )
+    except Exception as e:
+        logger.error(f"Generation failed: {e}")
+        return JSONResponse({"error": str(e)}, status_code=500)
+
+    return {
+        "model": DISPLAY_NAME,
+        "response": result["text"],
+        "done": True,
+        "eval_count": result["n_tokens"],
+        "eval_duration": int(result["elapsed"] * 1e9),
+    }
+
+
+@app.get("/health")
+async def health():
+    return {"status": "ok" if model is not None else "loading", "model": MODEL_PATH}