Upload app.py

app.py

@@ -366,23 +366,26 @@ def load_model():
         classifier_adapter_repo = "arvindsuresh-math/phi-4-error-binary-classifier"
         base_phi_model = "microsoft/Phi-4-mini-instruct"
 
-        DTYPE = torch.bfloat16
+        # T4 does fp16 (not bf16)
+        DTYPE = torch.float16
         quantization_config = BitsAndBytesConfig(
             load_in_4bit=True,
             bnb_4bit_quant_type="nf4",
-            bnb_4bit_compute_dtype=DTYPE
-            )
+            bnb_4bit_compute_dtype=DTYPE,
+        )
         classifier_backbone_base = AutoModelForCausalLM.from_pretrained(
             base_phi_model,
             quantization_config=quantization_config,
-            device_map="auto",
-            trust_remote_code=True,
-            )
+            device_map={"": 0},          # single-GPU
+            trust_remote_code=False,     # <-- avoid remote LossKwargs import
+            attn_implementation="sdpa",
+        )
 
         classifier_tokenizer = AutoTokenizer.from_pretrained(
             base_phi_model,
-            trust_remote_code=True
-            )
+            trust_remote_code=False      # <-- match above
+        )
+
         classifier_tokenizer.padding_side = "left"
         if classifier_tokenizer.pad_token is None:
             classifier_tokenizer.pad_token = classifier_tokenizer.eos_token
@@ -411,7 +414,6 @@ def load_model():
 def models_ready():
     return all([gemma_model, gemma_tokenizer, classifier_model, classifier_tokenizer])
 
-@spaces.GPU
 def analyze_single(math_question: str, proposed_solution: str, debug: bool = False):
     """
     Single (question, solution) classifier.
@@ -516,8 +518,6 @@ def analyze_single(math_question: str, proposed_solution: str, debug: bool = Fal
         }
 
 
-
-@spaces.GPU
 def classify_solution(question: str, solution: str):
     """
     Classify the math solution