Update app.py

app.py

@@ -1,99 +1,95 @@
-import spaces
+import time
+import torch
 import gradio as gr
-from transformers import AutoModelForCausalLM, AutoTokenizer
+from transformers import AutoTokenizer, AutoModelForCausalLM
 
-model_name = "daniel-dona/gemma-3-270m-it"
+MODEL_NAME = "google/gemma-3-270m-it"
 
-#pipe = pipeline("text-generation", model=model, device="cuda")
+# CPU optimizasyonları
+torch.set_num_threads(torch.get_num_threads())  # tüm çekirdekleri kullan
+torch.set_float32_matmul_precision("high")      # matmul hızını artır
 
-@spaces.GPU
-def respond(
-    message,
-    history: list[tuple[str, str]],
-    system_message,
-    max_tokens,
-    temperature,
-    top_p,
-):
-
-    messages = [{"role": "system", "content": system_message}]
-
-    print("Got:", message)
-
-    for val in history:
-        if val[0]:
-            messages.append({"role": "user", "content": val[0]})
-        if val[1]:
-            messages.append({"role": "assistant", "content": val[1]})
-
-    messages.append({"role": "user", "content": message})
-
-    """response = pipe(
-        messages,
+# Model/Tokenizer global yükleme
+tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_NAME,
+    torch_dtype=torch.float32,  # CPU'da float32
+    device_map=None
+)
+model.eval()
+
+# Kullanıcı bazlı KV cache
+sessions = {}  # {user_id: past_key_values}
+
+def build_prompt(message, history, system_message, max_ctx_tokens=1024):
+    msgs = [{"role": "system", "content": system_message}]
+    for u, a in history:
+        if u: msgs.append({"role": "user", "content": u})
+        if a: msgs.append({"role": "assistant", "content": a})
+    msgs.append({"role": "user", "content": message})
+
+    # Token bütçesi ile kırpma
+    while True:
+        text = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
+        if len(tokenizer(text, add_special_tokens=False).input_ids) <= max_ctx_tokens:
+            return text
+        # En eski user+assistant çiftini at
+        for i in range(1, len(msgs)):
+            if msgs[i]["role"] != "system":
+                del msgs[i:i+2]
+                break
+
+def respond(message, history, system_message, max_tokens, temperature, top_p):
+    user_id = "default"  # API bağlarsan burada kullanıcı ID'si ile değiştir
+    past = sessions.get(user_id)
+
+    if past is None:
+        # İlk mesaj → tüm prompt
+        text = build_prompt(message, history, system_message)
+        inputs = tokenizer([text], return_tensors="pt").to(model.device)
+    else:
+        # Sadece yeni mesajı encode et
+        inputs = tokenizer([message], return_tensors="pt").to(model.device)
+
+    do_sample = temperature > 0
+    gen_kwargs = dict(
         max_new_tokens=max_tokens,
-        temperature=temperature,
+        do_sample=do_sample,
         top_p=top_p,
-        return_full_text=False, 
-    )
-
-    generated_text = response[0]['generated_text']
-
-    yield generated_text"""
-
-    tokenizer = AutoTokenizer.from_pretrained(model_name)
-    model = AutoModelForCausalLM.from_pretrained(
-        model_name,
-        torch_dtype="auto",
-        device_map="auto"
+        temperature=temperature if do_sample else None,
+        use_cache=True,
+        past_key_values=past
     )
 
-    text = tokenizer.apply_chat_template(
-        messages,
-        tokenize=False,
-        add_generation_prompt=True
-    )
+    start_time = time.time()
+    with torch.inference_mode():
+        outputs = model.generate(**inputs, **{k: v for k, v in gen_kwargs.items() if v is not None},
+                                 return_dict_in_generate=True, output_scores=False)
+    end_time = time.time()
 
-    model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
+    # KV cache güncelle
+    sessions[user_id] = outputs.past_key_values
 
-    sample = True
+    # Yanıtı decode et
+    new_tokens = outputs.sequences[0][inputs["input_ids"].shape[1]:]
+    content = tokenizer.decode(new_tokens, skip_special_tokens=True).strip("\n")
 
-    if temperature == 0:
-        sample = False
+    # T/S hesapla
+    token_count = len(new_tokens)
+    elapsed = end_time - start_time
+    tps = token_count / elapsed if elapsed > 0 else 0
 
-    # conduct text completion
-    generated_ids = model.generate(
-        **model_inputs,
-        max_new_tokens=max_tokens,
-        do_sample=sample,
-        top_p=top_p,
-        temperature=temperature
-    )
-    output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 
+    return f"{content}\n\n⚡ **Hız:** {tps:.2f} token/sn"
 
-    content = tokenizer.decode(output_ids, skip_special_tokens=True).strip("\n")
-
-    return content
-
-
-"""
-For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
-"""
 demo = gr.ChatInterface(
     respond,
     additional_inputs=[
         gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
-        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
+        gr.Slider(minimum=1, maximum=2048, value=256, step=1, label="Max new tokens"),
         gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
-        gr.Slider(
-            minimum=0.1,
-            maximum=1.0,
-            value=0.95,
-            step=0.05,
-            label="Top-p (nucleus sampling)",
-        ),
+        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p"),
     ],
 )
 
-
 if __name__ == "__main__":
     demo.launch()