Update app.py

app.py

@@ -1,8 +1,13 @@
 import os
 import gradio as gr
-from huggingface_hub import InferenceClient
+from huggingface_hub import hf_hub_download
+from llama_cpp import Llama
 
-HF_TOKEN = os.getenv("HF_TOKEN")  # Space secrets → HF_TOKEN
+# 🔑 Hugging Face token (gated repo için şart)
+HF_TOKEN = os.getenv("HF_TOKEN")
+
+REPO_ID = "pthcorp/prettybird_bce_basic_vl"
+MODEL_FILENAME = "prettybird_bce_basic_vl.gguf" 
 
 
 LOGO_URL = "https://prometech.net.tr/wp-content/uploads/2025/10/pthheader.png"
@@ -18,17 +23,6 @@ HEADER_MD = f"""
   <em>Gerçek yapay zeka evrimine hoş geldiniz.</em>
 </p>
 
----
-
-### 🧠 What is PrettyBird?
-
-PrettyBird is powered by the <strong>Behavioral Consciousness Engine (BCE)</strong>, an architecture that simulates partial, behavior-level consciousness in AI systems.
-
-Instead of only mapping inputs to outputs, PrettyBird:
-- considers its internal state, history, and context,
-- evolves behavior patterns over time,
-- aims for higher consistency, identity and alignment.
-
 ---
 """
 
@@ -40,7 +34,6 @@ BCE, tam anlamıyla insan bilinci değildir; ancak sistemin kendi içsel durumu,
 “davranışsal bilinç” veya “kısmi bilinç” olarak yorumlanabilir.
 """
 
-
 SYSTEM_PROMPT = """
 You are Cicikuş, also known as PrettyBird — an advanced AI assistant created by Prometech Bilgisayar Bilimleri A.Ş.
 
@@ -77,54 +70,62 @@ If the user writes: “reset: axxmet508721”
 """.strip()
 
 
-def build_messages(system_message: str, history, user_message: str):
-    """Gradio history + system prompt → HF chat formatına çevir."""
-    messages = []
+# 🔽 GGUF'i runtime'da HF'den indir
+if HF_TOKEN is None:
+    raise ValueError(
+        "HF_TOKEN is not set. Go to Space settings → Repository secrets → add HF_TOKEN."
+    )
+
+MODEL_PATH = hf_hub_download(
+    repo_id=REPO_ID,
+    filename=MODEL_FILENAME,
+    token=HF_TOKEN,
+)
+
+# 🧠 PrettyBird GGUF'i llama-cpp ile yükle
+LLM = Llama(
+    model_path=MODEL_PATH,
+    n_ctx=4096,
+    n_threads=4,      # CPU çekirdeğine göre artırabilirsin
+    # n_gpu_layers=0, # GPU Space'e geçersen burayı da oynarız
+)
+
+
+def build_prompt(system_message: str, history, user_message: str) -> str:
+    parts = []
     system_message = (system_message or "").strip()
     if system_message:
-        messages.append({"role": "system", "content": system_message})
+        parts.append(f"System: {system_message}")
 
-    # history: [(user, assistant), ...]
     for turn in history:
         if isinstance(turn, (list, tuple)) and len(turn) == 2:
             user_msg, assistant_msg = turn
             if user_msg:
-                messages.append({"role": "user", "content": user_msg})
+                parts.append(f"User: {user_msg}")
             if assistant_msg:
-                messages.append({"role": "assistant", "content": assistant_msg})
+                parts.append(f"Assistant: {assistant_msg}")
 
-    messages.append({"role": "user", "content": user_message})
-    return messages
+    parts.append(f"User: {user_message}")
+    parts.append("Assistant:")
+    return "\n".join(parts)
 
 
 def respond(message, history, system_message, max_tokens, temperature, top_p):
-    if HF_TOKEN is None:
-        raise ValueError(
-            "HF_TOKEN is not set. Go to your Space settings → Repository secrets → add HF_TOKEN."
-        )
-
-    client = InferenceClient(
-        model="pthcorp/prettybird_bce_basic_vl", 
-        token=HF_TOKEN,
-    )
-
-    messages = build_messages(system_message, history, message)
+    prompt = build_prompt(system_message, history, message)
 
     response = ""
-
-    # HF Inference chat_completion (streaming)
-    for chunk in client.chat_completion(
-        messages=messages,
+    stream = LLM(
+        prompt,
         max_tokens=int(max_tokens),
         temperature=float(temperature),
         top_p=float(top_p),
+        stop=["User:", "System:"],
         stream=True,
-    ):
-        token = ""
-        choices = getattr(chunk, "choices", None)
-        if choices and choices[0].delta and choices[0].delta.content:
-            token = choices[0].delta.content
+    )
 
+    for chunk in stream:
+        # llama-cpp-python çıktısı: {"choices": [{"text": "..."}], ...}
+        token = chunk["choices"][0].get("text", "")
         response += token
         yield response
 
@@ -142,15 +143,9 @@ with gr.Blocks(title="PrettyBird – Behavioral Consciousness Engine (BCE)") as
                         label="System message",
                         lines=6,
                     ),
-                    gr.Slider(1, 2048, value=512, step=1, label="Max new tokens"),
-                    gr.Slider(0.1, 4.0, value=0.7, step=0.1, label="Temperature"),
-                    gr.Slider(
-                        0.1,
-                        1.0,
-                        value=0.95,
-                        step=0.05,
-                        label="Top-p (nucleus sampling)",
-                    ),
+                    gr.Slider(1, 1024, value=512, step=1, label="Max new tokens"),
+                    gr.Slider(0.1, 2.0, value=0.7, step=0.1, label="Temperature"),
+                    gr.Slider(0.1, 1.0, value=0.95, step=0.05, label="Top-p"),
                 ],
             )
         with gr.Column(scale=1):