Do not use safetensors

app.py

@@ -38,11 +38,12 @@ print("Loaded vardecoder model successfully.")
 logger.info("Loading fielddecoder model...")
 
 fielddecoder_model = None
-#fielddecoder_model = AutoModelForCausalLM.from_pretrained(
-#    "ejschwartz/resym-fielddecoder", 
-#    torch_dtype=torch.bfloat16,
-#)
-#logger.info("Successfully loaded fielddecoder model")
+fielddecoder_model = AutoModelForCausalLM.from_pretrained(
+   "ejschwartz/resym-fielddecoder", 
+   torch_dtype=torch.bfloat16,
+   use_safetensors=False
+)
+logger.info("Successfully loaded fielddecoder model")
 
 make_gradio_client = lambda: Client("https://ejschwartz-resym-field-helper.hf.space/")
 
@@ -104,7 +105,7 @@ def infer(code):
 
     print(f"Prompt:\n{repr(var_prompt)}")
 
-    var_input_ids = tokenizer.encode(var_prompt, return_tensors="pt").cuda()[
+    var_input_ids = tokenizer.encode(var_prompt, return_tensors="pt").to(vardecoder_model.device)[
         :, : MAX_CONTEXT_LENGTH - MAX_NEW_TOKENS
     ]
     var_output = vardecoder_model.generate(
@@ -131,7 +132,7 @@ def infer(code):
     if len(fields) == 0:
         field_output = "Failed to parse fields" if field_prompt_result is None else "No fields"
     else:
-        field_input_ids = tokenizer.encode(field_prompt_result, return_tensors="pt").cuda()[
+        field_input_ids = tokenizer.encode(field_prompt_result, return_tensors="pt").to(fielddecoder_model.device)[
             :, : MAX_CONTEXT_LENGTH - MAX_NEW_TOKENS
         ]