Create app.py

app.py

@@ -0,0 +1,107 @@
+import gradio as gr
+import torch
+import re
+from transformers import AutoModelForCausalLM, PreTrainedTokenizerFast
+
+# ==============================================================================
+# 1. MODEL & TOKENIZER INITIALIZATION
+# ==============================================================================
+# Menggunakan repository organisasi barumu
+repo_id = "VoidSpac3s/Lumina-Medium-390M" 
+
+print("Loading Tokenizer...")
+tokenizer = PreTrainedTokenizerFast.from_pretrained(repo_id)
+
+print("Loading Lumina-Medium-390M to CPU...")
+# Tetap menggunakan float32 karena ini yang paling stabil untuk HF Spaces CPU Basic
+model = AutoModelForCausalLM.from_pretrained(
+    repo_id,
+    dtype=torch.float32 
+)
+model.to("cpu")
+
+# ==============================================================================
+# 2. INFERENCE FUNCTION
+# ==============================================================================
+def generate_prompt(raw_prompt, temperature, top_k, top_p, rep_penalty, max_tokens):
+    if not raw_prompt.strip():
+        return "⚠️ Please enter a base prompt first."
+    
+    prompt_full = f"<|im_start|>user\nEnhance this prompt: {raw_prompt}<|im_end|>\n<|im_start|>assistant\n"
+    
+    input_ids = tokenizer(prompt_full, return_tensors="pt").input_ids.to("cpu")
+    input_length = input_ids.shape[1] 
+    
+    with torch.no_grad():
+        output = model.generate(
+            input_ids,
+            max_new_tokens=max_tokens,
+            temperature=temperature,      
+            top_k=int(top_k),
+            top_p=top_p,
+            repetition_penalty=rep_penalty, 
+            do_sample=True,
+            eos_token_id=tokenizer.eos_token_id,
+            pad_token_id=tokenizer.pad_token_id
+        )
+        
+    # Hanya ambil token yang baru saja di-generate
+    generated_tokens = output[0][input_length:]
+    
+    enhanced_prompt = tokenizer.decode(generated_tokens, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+    
+    # Fallback pembersihan simbol Tokenizer
+    enhanced_prompt = enhanced_prompt.replace("Ġ", " ").replace("Ċ", "\n").replace("<|im_end|>", "").strip()
+    
+    # FIX: Membersihkan spasi aneh sebelum tanda baca (contoh: "eyes , highly" menjadi "eyes, highly")
+    enhanced_prompt = re.sub(r'\s+([.,!?])', r'\1', enhanced_prompt)
+    
+    return enhanced_prompt
+
+# ==============================================================================
+# 3. GRADIO UI
+# ==============================================================================
+with gr.Blocks(theme=gr.themes.Soft(primary_hue="indigo")) as demo:
+    gr.Markdown(
+        """
+        # ✨ Lumina-Medium Prompt Enhancer
+        **Lumina-Medium-390M** is a custom lightweight language model (~390M Parameters) trained specifically on 2.4 Million data points to enhance base ideas into highly detailed, context-aware image generation prompts.
+        """
+    )
+    
+    with gr.Row():
+        with gr.Column(scale=2):
+            input_text = gr.Textbox(
+                label="Base Prompt", 
+                placeholder="Example: A futuristic cyberpunk samurai cat", 
+                lines=2
+            )
+            generate_btn = gr.Button("🚀 Generate Prompt", variant="primary")
+            
+        with gr.Column(scale=3):
+            output_text = gr.Textbox(
+                label="Lumina Enhanced Prompt", 
+                lines=6, 
+                interactive=False
+            )
+            
+    # Advanced Settings
+    with gr.Accordion("⚙️ Advanced Settings", open=False):
+        with gr.Row():
+            temperature = gr.Slider(minimum=0.1, maximum=1.5, value=0.7, step=0.1, label="Temperature (Creativity)")
+            top_k = gr.Slider(minimum=1, maximum=100, value=50, step=1, label="Top-K")
+            top_p = gr.Slider(minimum=0.1, maximum=1.0, value=0.9, step=0.05, label="Top-P")
+        with gr.Row():
+            rep_penalty = gr.Slider(minimum=1.0, maximum=2.0, value=1.1, step=0.05, label="Repetition Penalty")
+            max_tokens = gr.Slider(minimum=50, maximum=256, value=150, step=10, label="Max New Tokens")
+
+    # Button Logic
+    generate_btn.click(
+        fn=generate_prompt,
+        inputs=[input_text, temperature, top_k, top_p, rep_penalty, max_tokens],
+        outputs=output_text
+    )
+
+# Run the app
+if __name__ == "__main__":
+    demo.launch()