app.py

import torch
from fastapi import FastAPI
from pydantic import BaseModel
from huggingface_hub import list_repo_files

MODEL_ID = "peterkirby/modernbert-large-pan2020-authorship-verification"

# Определяем тип модели: bi-encoder (sentence-transformers) или pair classifier
IS_BIENCODER = "modules.json" in list_repo_files(MODEL_ID)

if IS_BIENCODER:
    from sentence_transformers import SentenceTransformer, util
    model = SentenceTransformer(MODEL_ID)
else:
    from transformers import AutoTokenizer, AutoModelForSequenceClassification
    tok = AutoTokenizer.from_pretrained(MODEL_ID)
    model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID).eval()

class Req(BaseModel):
    text: str
    text_pair: str

app = FastAPI()

@app.get("/")
def health():
    return {"ok": True, "mode": "bi-encoder" if IS_BIENCODER else "pair-classifier"}

@app.post("/verify")
def verify(r: Req):
    if IS_BIENCODER:
        e = model.encode([r.text, r.text_pair], convert_to_tensor=True,
                         normalize_embeddings=True)
        p = (float(util.cos_sim(e[0], e[1])) + 1) / 2
    else:
        enc = tok(r.text, r.text_pair, return_tensors="pt",
                  truncation=True, max_length=512)
        with torch.no_grad():
            logits = model(**enc).logits[0]
        p = float(torch.sigmoid(logits[0])) if logits.numel() == 1 \
            else float(torch.softmax(logits, -1)[1])
    # тот же формат, что возвращал HF API, чтобы остальной код не менять
    return [[{"label": "LABEL_1", "score": p},
             {"label": "LABEL_0", "score": 1 - p}]]