Create app.py

app.py

@@ -0,0 +1,63 @@
+from fastapi import FastAPI
+from pydantic import BaseModel
+import torch, re, asyncio, aiohttp, os
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+
+MODEL_ID = os.getenv("MODEL_ID", "ai-forever/mGPT-1.3B-persian")
+device = "cuda" if torch.cuda.is_available() else "cpu"
+dtype = torch.float16 if device == "cuda" else torch.float32
+
+
+# کم‌مصرف روی CPU
+torch.set_num_threads(1)
+
+
+app = FastAPI()
+
+
+tok = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=True)
+model = AutoModelForCausalLM.from_pretrained(
+MODEL_ID,
+torch_dtype=dtype,
+low_cpu_mem_usage=True
+).to(device).eval()
+
+
+class Req(BaseModel):
+prompt: str
+max_tokens: int = 160
+system: str = "تو یه دستیار فارسی خودمونی و سریع هستی؛ جواب‌ها کوتاه، رک و بامزه (۱–۲ جمله)."
+temperature: float = 0.65
+
+
+@app.get("/health")
+def health():
+return {"ok": True}
+
+
+@app.get("/")
+def root():
+return {"ok": True, "use": "POST /generate"}
+
+
+def _clean(txt: str) -> str:
+txt = txt.replace("[دستیار]:", "").replace("[سیستم]:", "").replace("[کاربر]:", "")
+txt = re.sub(r"\[[^\]\n]{0,12}\]:", "", txt).strip()
+parts = re.split(r"(?<=[.!؟?])\s+", txt)
+short = " ".join(parts[:2]).strip() or txt
+return short[:220]
+
+
+@app.post("/generate")
+def generate(r: Req):
+sys = (r.system or "")[:400]
+user = r.prompt[:900]
+text_in = f"[سیستم]: {sys}\n[کاربر]: {user}\n[دستیار]:"
+inputs = tok(text_in, return_tensors="pt").to(device)
+with torch.no_grad():
+out = model.generate(
+**inputs,
+max_new_tokens=min(200, r.max_tokens),
+do_sample=True,
+asyncio.create_task(_keepalive())