Upload model.py with huggingface_hub

model.py

@@ -1,13 +1,13 @@
 """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.
+The actual HF architecture lives in modeling_lemondecider.py/configuration_lemondecider.py.
 """
 import argparse
 import json
 from pathlib import Path
 import torch
-from transformers import AutoModel, AutoProcessor
+from transformers import AutoModel
 
 if Path(__file__).with_name('decision_utils.py').is_file():
     from decision_utils import predict_batch, answer_codes
@@ -22,11 +22,17 @@ def load_model(checkpoint, device='cuda:0'):
     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 Path(__file__).with_name('decision_utils.py').is_file():
+        from decision_utils import load_processor
+    else:
+        from modeling.decision_utils import load_processor
+    processor = load_processor(checkpoint)
+    codes, ids = answer_codes(processor.tokenizer, model.config.max_options,
+                              getattr(model.config, 'code_scheme', 'letters-v1'))
     if codes != model.config.codes or ids != model.config.token_ids:
         raise ValueError('Checkpoint code/tokenizer mapping mismatch')
+    if tuple(model.readout.weight.shape) != (len(codes), model.config.text_config.hidden_size):
+        raise ValueError('Checkpoint readout/code-table shape mismatch')
     return model, processor