New: Load prediction model

app.py

@@ -1,8 +1,10 @@
+import os
 import gradio as gr
 import fasttext
 import html
 import numpy as np
 import types
+import torch
 from huggingface_hub import hf_hub_download
 from safetensors.torch import load_file
 from transformers import AutoTokenizer
@@ -19,6 +21,37 @@ from lib.bert_regressor_utils import (
 
 ### Stettings ####################################################################
 
+MODEL_BASE = "microsoft/deberta-v3-base"
+REPO_ID    = "ziem-io/deberta_flavour_regressor_multi_head"
+FILENAME   = "deberta_flavour_regressor_multi_head_20250914_1020.safetensors"
+
+# (optional) falls das Model-Repo privat ist:
+HF_TOKEN = os.getenv("HF_TOKEN")  # in Space-Secrets hinterlegen
+
+##################################################################################
+
+# --- Download Weights ---
+weights_path = hf_hub_download(
+    repo_id=REPO_ID, 
+    filename=FILENAME, 
+    token=HF_TOKEN
+)
+
+# --- Tokenizer (SentencePiece!) ---
+tokenizer_flavours = AutoTokenizer.from_pretrained(
+    MODEL_BASE, 
+    use_fast=False
+)
+
+model_flavours = BertMultiHeadRegressor(
+    pretrained_model_name=MODEL_BASE
+)
+state = load_file(weights_path)                    # safetensors -> dict[str, Tensor]
+_ = model_flavours.load_state_dict(state, strict=False)  # strict=True wenn Keys exakt passen
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+model_flavours.to(device).eval()
+
 ##################################################################################
 
 # offizielles Mirror-Repo mit lid.176.*