Use separate var name

app.py

@@ -14,7 +14,6 @@ huggingface_hub.login(token=hf_key)
 
 tokenizer = AutoTokenizer.from_pretrained(
     "bigcode/starcoderbase-3b"
-    # , use_auth_token=hf_key
 )
 vardecoder_model = AutoModelForCausalLM.from_pretrained(
     "ejschwartz/resym-vardecoder", torch_dtype=torch.bfloat16, device_map="auto"
@@ -38,10 +37,10 @@ example = """{
 
 
 @spaces.GPU
-def infer(input):
+def infer(var_name, input):
     line = json.loads(input)
     first_token = line["output"].split(":")[0]
-    prompt = line["input"] + first_token + ":"
+    prompt = line["input"] + var_name + ":"
 
     input_ids = tokenizer.encode(prompt, return_tensors="pt").cuda()[:, : 8192 - 1024]
     output = vardecoder_model.generate(