Update app.py

app.py

@@ -36,9 +36,6 @@ humanize_pipe = None
 
 # Load NLP models
 try:
-    # Initialize with a flag to ensure loading only happens once
-    model_loaded = False
-    
     # Load spaCy model
     nlp = spacy.load("en_core_web_sm")
     
@@ -55,7 +52,7 @@ except Exception as e:
 def get_humanize_pipeline():
     """
     Lazy-load the humanization pipeline on first use.
-    Ensures it runs on CPU with limited memory settings.
+    Uses standard settings that don't require accelerate.
     """
     global humanize_pipe
     if humanize_pipe is None:
@@ -65,36 +62,30 @@ def get_humanize_pipeline():
             # Force CPU usage
             device = torch.device("cpu")
             
-            # Set low memory footprint
-            model_kwargs = {
-                "low_cpu_mem_usage": True,
-                "device_map": "cpu"
-            }
-            
-            # Load model with specific settings for resource-constrained environments
+            # Load model with basic settings (no accelerate needed)
             model = AutoModelForSeq2SeqLM.from_pretrained(
                 "danibor/flan-t5-base-humanizer", 
-                **model_kwargs,
                 torch_dtype=torch.float32  # Use float32 instead of float16 for CPU
             )
             tokenizer = AutoTokenizer.from_pretrained("danibor/flan-t5-base-humanizer")
             
-            # Create pipeline with optimized settings
+            # Create pipeline with basic settings
             humanize_pipe = pipeline(
                 "text2text-generation", 
                 model=model, 
                 tokenizer=tokenizer,
-                device=device,  # Explicitly specify CPU
-                framework="pt"
+                device=device  # Explicitly specify CPU
             )
             
             print("Humanizer model loaded successfully!")
+            return humanize_pipe
         except Exception as e:
             print(f"Error loading humanizer model: {e}")
-            # Return a simple function that just returns the input as fallback
-            def fallback_humanize(text, **kwargs):
-                return [{"generated_text": f"FALLBACK: {text} (Model loading failed: {str(e)})"}]
-            humanize_pipe = fallback_humanize
+            # Create a simple pipeline-like function that just returns the input
+            def simple_pipeline(text, **kwargs):
+                return [{"generated_text": f"Could not process: {text} (Model failed to load)"}]
+            humanize_pipe = simple_pipeline
+            return humanize_pipe
     
     return humanize_pipe
 
@@ -127,13 +118,11 @@ async def humanize_text(request: TextRequest):
         # Get or initialize the pipeline
         pipeline = get_humanize_pipeline()
         
-        # Generate humanized text with memory-conscious settings
+        # Generate humanized text with basic settings
         result = pipeline(
             input_text, 
             max_length=min(500, len(input_text) * 2),  # Limit max length
-            do_sample=True,
-            num_return_sequences=1,
-            batch_size=1  # Small batch size for memory constraints
+            do_sample=True
         )
         
         humanized_text = result[0]['generated_text']