Upload model.py with huggingface_hub

model.py

@@ -0,0 +1,43 @@
+"""Standalone local decision inference (not an AIStudio Maya canvas component).
+
+Inside the exported checkpoint this file can be run directly with --checkpoint .
+The actual HF architecture lives in modeling_autojev.py/configuration_autojev.py.
+"""
+import argparse
+import json
+from pathlib import Path
+import torch
+from transformers import AutoModel, AutoProcessor
+
+if Path(__file__).with_name('decision_utils.py').is_file():
+    from decision_utils import predict_batch, answer_codes
+else:
+    from modeling.decision_utils import predict_batch, answer_codes
+
+
+def load_model(checkpoint, device='cuda:0'):
+    torch.set_num_threads(8)
+    if torch.cuda.is_available():
+        torch.backends.cuda.enable_cudnn_sdp(False)
+    model = AutoModel.from_pretrained(str(checkpoint), trust_remote_code=True,
+        local_files_only=True, dtype=torch.bfloat16, device_map=device,
+        attn_implementation='sdpa').eval()
+    processor = AutoProcessor.from_pretrained(str(checkpoint), local_files_only=True, trust_remote_code=True)
+    processor.tokenizer.padding_side = 'left'
+    codes, ids = answer_codes(processor.tokenizer)
+    if codes != model.config.codes or ids != model.config.token_ids:
+        raise ValueError('Checkpoint code/tokenizer mapping mismatch')
+    return model, processor
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--checkpoint', default=str(Path(__file__).resolve().parent))
+    parser.add_argument('--device', default='cuda:0')
+    parser.add_argument('--input', help='JSON file containing a list of decision inputs')
+    args = parser.parse_args()
+    model, processor = load_model(args.checkpoint, args.device)
+    rows = json.loads(Path(args.input).read_text()) if args.input else [{
+        'state': '我想申请退款。', 'question': {'type': 'choice', 'instructions': '判断主要诉求',
+        'criteria': {'refund': '退款相关', 'delivery': '物流相关', 'account': '账户相关'}}}]
+    print(json.dumps(predict_batch(model, processor, rows), ensure_ascii=False, indent=2))