Update app.py

app.py

@@ -1,7 +1,8 @@
 import gradio as gr
 import torch
 import spaces
-from transformers import AutoModelForCausalLM, AutoTokenizer
+from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
+from threading import Thread
 
 model_name = "Pyefuri/Qwen2.5-3B-Bahasa-Biak-Final"
 
@@ -9,7 +10,6 @@ print("Memuat tokenizer...")
 tokenizer = AutoTokenizer.from_pretrained(model_name)
 
 print("Memuat model Bahasa Biak...")
-# Mengembalikan argumen ke 'dtype' sesuai warning log transformers Anda agar tidak deprecated
 model = AutoModelForCausalLM.from_pretrained(
     model_name,
     dtype=torch.float16,
@@ -48,10 +48,13 @@ def prediksi_terjemahan(message, history):
     )
     
     inputs = tokenizer([input_text], return_tensors="pt").to(model.device)
-    input_len = inputs.input_ids.shape
+    
+    # SOLUSI PERMANEN: Menggunakan Streamer bawaan Hugging Face untuk menangani teks mengalir
+    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
     
     generation_kwargs = dict(
         **inputs,
+        streamer=streamer,  # Menyuntikkan fungsi pengalir teks otomatis
         max_new_tokens=128,
         temperature=0.4,
         top_p=0.85,
@@ -61,51 +64,37 @@ def prediksi_terjemahan(message, history):
         pad_token_id=tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
     )
     
-    with torch.no_grad():
-        generated_tokens = inputs.input_ids
-        for _ in range(generation_kwargs["max_new_tokens"]):
-            outputs = model(**{"input_ids": generated_tokens})
-            next_token_logits = outputs.logits[:, -1, :]
-            
-            if generation_kwargs["do_sample"]:
-                next_token_logits = next_token_logits / generation_kwargs["temperature"]
-                probs = torch.softmax(next_token_logits, dim=-1)
-                next_token = torch.multinomial(probs, num_samples=1)
-            else:
-                next_token = torch.argmax(next_token_logits, dim=-1, keepdim=True)
-                
-            generated_tokens = torch.cat([generated_tokens, next_token], dim=-1)
-            text_terjemahan = tokenizer.decode(generated_tokens[input_len:], skip_special_tokens=True)
-            
-            yield text_terjemahan.strip()
-            
-            if next_token.item() == generation_kwargs["eos_token_id"]:
-                break
+    # Menjalankan proses generasi AI di latar belakang (Thread) agar UI web tidak macet
+    thread = Thread(target=model.generate, kwargs=generation_kwargs)
+    thread.start()
+    
+    # Menangkap kata demi kata yang keluar dari streamer dan mengirimkannya langsung ke UI
+    text_terjemahan = ""
+    for new_text in streamer:
+        text_terjemahan += new_text
+        yield text_terjemahan.strip()
 
 # ==========================================
 #         STRUKTUR DESAIN CUSTOM CSS
 # ==========================================
 desain_kustom_css = """
 #wadah-lebar-chat {
-    max-width: 800px !important;   /* Membatasi lebar agar pas seperti ChatGPT */
+    max-width: 800px !important;   /* Membatasi lebar agar pas di tengah seperti ChatGPT */
     margin: 0 auto !important;      /* Menjaga posisi tetap presisi di tengah layar */
     float: none !important;
 }
 """
 
-# Membuka Blocks tanpa menyelipkan theme/css langsung di sini agar lolos dari Warning Gradio
 with gr.Blocks() as demo:
     gr.Markdown("# 🤖 AI Bahasa Indonesia-Biak")
     
-    # Solusi: Membungkus ChatInterface ke dalam gr.Column yang mendukung penuh elem_id
     with gr.Column(elem_id="wadah-lebar-chat"):
         gr.ChatInterface(
             fn=prediksi_terjemahan,
-            
+           
         )
 
 if __name__ == "__main__":
-    # PERBAIKAN: Memindahkan pengaturan tema dan CSS ke fungsi .launch() sesuai standarisasi sistem Gradio
     demo.launch(
         theme=gr.themes.Soft(),
         css=desain_kustom_css