Use smaller model and simpler config for Zero GPU compatibility

app.py

@@ -75,10 +75,9 @@ You are a mathematics expert. Solve grade school math problems step by step.
         progress(0.3, desc="Loading model and tokenizer...")
         output_log.append(f"\n🤖 Loading {model_name}...")
         
-        # Use smaller model for demo
-        if "7B" in model_name:
-            model_name = "Qwen/Qwen2.5-1.5B"  # Use smaller model for Zero GPU demo
-            output_log.append("   Note: Using 1.5B model for Zero GPU compatibility")
+        # Always use smaller model for Zero GPU demo
+        model_name = "Qwen/Qwen2.5-0.5B"  # Use smallest model for Zero GPU
+        output_log.append("   Note: Using 0.5B model for Zero GPU compatibility")
         
         tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
         if tokenizer.pad_token is None:
@@ -138,14 +137,15 @@ You are a mathematics expert. Solve grade school math problems step by step.
             output_dir="./qwen-promptwizard-zerogpu",
             num_train_epochs=num_epochs,
             per_device_train_batch_size=batch_size,
-            gradient_accumulation_steps=4,
-            warmup_steps=50,
-            logging_steps=10,
+            gradient_accumulation_steps=2,  # Reduced for Zero GPU
+            warmup_steps=10,  # Reduced warmup
+            logging_steps=5,
             save_strategy="no",  # Don't save during demo
-            fp16=True,
-            gradient_checkpointing=True,
+            fp16=device == "cuda",  # Only use fp16 on GPU
+            gradient_checkpointing=False,  # Disable for simplicity
             optim="adamw_torch",
             learning_rate=learning_rate,
+            max_steps=50,  # Limit steps for demo
         )
         
         # Create trainer
@@ -209,11 +209,11 @@ def create_interface():
                 
                 model_name = gr.Dropdown(
                     choices=[
+                        "Qwen/Qwen2.5-0.5B",
                         "Qwen/Qwen2.5-1.5B",
-                        "Qwen/Qwen2.5-7B",
                     ],
-                    value="Qwen/Qwen2.5-1.5B",
-                    label="Model (1.5B recommended for Zero GPU)"
+                    value="Qwen/Qwen2.5-0.5B",
+                    label="Model (0.5B works best for Zero GPU)"
                 )
                 
                 num_epochs = gr.Slider(