add custom layernorm patch, low mem usage

tasks/text.py

@@ -50,7 +50,7 @@ class TextClassifier:
                 # Initialize tokenizer
                 self.tokenizer = AutoTokenizer.from_pretrained(
                     model_name,
-                    model_max_length=512,
+                    model_max_length=2048,
                     padding_side='right',
                     truncation_side='right'
                 )
@@ -64,7 +64,8 @@ class TextClassifier:
                     self.model = AutoModelForSequenceClassification.from_pretrained(
                         model_name,
                         config=self.config,
-                        ignore_mismatched_sizes=True
+                        ignore_mismatched_sizes=True,
+                        low_cpu_mem_usage=True
                     )
                 finally:
                     # Restore original LayerNorm
@@ -91,7 +92,7 @@ class TextClassifier:
                 batch,
                 return_tensors="pt",
                 truncation=True,
-                max_length=512,
+                max_length=2048,
                 padding='max_length'
             ).to(self.device)
             
@@ -146,7 +147,7 @@ async def evaluate_text(request: TextEvaluationRequest):
         classifier = TextClassifier()
 
         # Prepare batches
-        batch_size = 32
+        batch_size = 24
         quotes = test_dataset["quote"]
         num_batches = len(quotes) // batch_size + (1 if len(quotes) % batch_size != 0 else 0)
         batches = [