Update handler.py

handler.py

@@ -1,12 +1,13 @@
 # -*- coding: utf-8 -*-
 """
-PULSE ECG Handler — Demo Parity Mode
+PULSE ECG Handler — Demo Parity + Style Hint
 - Demo app.py ile aynı üretim ayarları:
   do_sample=True, temperature=0.05, top_p=1.0, max_new_tokens=4096
 - Stopping: konuşma ayırıcıda (conv.sep/sep2) güvenli token-eşleşmeli kriter
 - Görsel tensörü: .half() ve model cihazında
 - Streamer: TextIteratorStreamer (demo gibi), thread ile generate
 - Seed/deterministic KAPALI (göndermezseniz); demo gibi stokastik
+- STYLE_HINT: demo üslubuna (narratif + sonda tek satır structured impression) yaklaşmak için
 """
 
 import os
@@ -80,6 +81,13 @@ context_len = None
 args = None
 model_initialized = False
 
+# ====== Demo üslubuna yönlendiren stil ipucu ======
+STYLE_HINT = (
+    "Write a concise diagnostic narrative as in a cardiology read: "
+    "use 2–3 short paragraphs describing rhythm, rate, axis, chamber enlargement, conduction, QRS, ST–T, QT; "
+    "then finish with a single final line starting exactly with 'Structured clinical impression:'. "
+    "Do not include recommendations, prognosis, follow-up, or risk counseling. No emojis or bullet points."
+)
 
 # ===================== Utilities =====================
 
@@ -140,7 +148,6 @@ def _wrap_image_token_if_needed(model_cfg) -> bool:
     except Exception:
         return False
 
-
 # ====== Güvenli Stop Kriteri (demo eşleniği) ======
 class SafeKeywordsStoppingCriteria(StoppingCriteria):
     """
@@ -153,7 +160,6 @@ class SafeKeywordsStoppingCriteria(StoppingCriteria):
         self.kw_ids = tok  # shape: (n,)
 
     def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
-        # input_ids: (bsz, seq_len)
         if input_ids is None or input_ids.shape[0] == 0:
             return False
         out = input_ids[0]  # assume bsz=1
@@ -161,18 +167,15 @@ class SafeKeywordsStoppingCriteria(StoppingCriteria):
         if out.shape[0] < n:
             return False
         tail = out[-n:]
-        # cihaz hizası
         kw = self.kw_ids.to(tail.device)
         return torch.equal(tail, kw)
 
-
 # ===================== Core Generation =====================
 
 def _build_prompt_and_ids(chatbot, user_text: str, device: torch.device):
     # demo gibi: <image> + text (IM_START/END gerekiyorsa sar)
     use_wrap = _wrap_image_token_if_needed(chatbot.model.config)
     if use_wrap:
-        # <im_start><image><im_end>\n + user text
         inp = f"{DEFAULT_IM_START_TOKEN}{DEFAULT_IMAGE_TOKEN}{DEFAULT_IM_END_TOKEN}\n{user_text}"
     else:
         inp = f"{DEFAULT_IMAGE_TOKEN}\n{user_text}"
@@ -195,7 +198,6 @@ def generate_response(
     max_new_tokens: Optional[int] = None,
     conv_mode_override: Optional[str] = None,
     repetition_penalty: Optional[float] = None,  # demo'da yok; verilirse 1.0 yaparız
-    # NOT: no_repeat_ngram_size / min_new_tokens / custom_stop KULLANMIYORUZ → demo-parite
     det_seed: Optional[int] = None,              # seed gönderilmezse stokastik (demo gibi)
 ):
     if not (LLAVA_AVAILABLE and TRANSFORMERS_AVAILABLE):
@@ -237,30 +239,31 @@ def generate_response(
 
     # Cihaz/dtype
     device = next(chatbot.model.parameters()).device
-    # demo half: .half() kullanacağız
-    dtype = torch.float16
+    dtype = torch.float16  # demo: half
 
     # Görüntü ön-işleme → tensör
     try:
         processed = process_images([pil_img], chatbot.image_processor, chatbot.model.config)
-        # LLaVA genelde list döndürür
         if isinstance(processed, (list, tuple)) and len(processed) > 0:
             image_tensor = processed[0]
         elif isinstance(processed, torch.Tensor):
-            image_tensor = processed[0] if processed.ndim == 4 else processed  # güvenlik
+            image_tensor = processed[0] if processed.ndim == 4 else processed
         else:
             return {"error": "Image processing returned empty"}
         if image_tensor.ndim == 3:
             image_tensor = image_tensor.unsqueeze(0)  # (1,C,H,W)
-        # demo: half + device
-        image_tensor = image_tensor.to(device=device, dtype=dtype)
+        image_tensor = image_tensor.to(device=device, dtype=dtype)  # demo: half + device
     except Exception as e:
         return {"error": f"Image processing failed: {e}"}
 
+    # --------- STIL İPUCU EKLEME ---------
+    message_text = (message_text or "").strip() + "\n\n" + STYLE_HINT
+    # -------------------------------------
+
     # Prompt & input ids
     _, input_ids = _build_prompt_and_ids(chatbot, message_text, device)
 
-    # Stop string from conv
+    # Stop string (conv separator) → güvenli kriter
     stop_str = chatbot.conversation.sep if chatbot.conversation.sep_style != SeparatorStyle.TWO else chatbot.conversation.sep2
     stopping = SafeKeywordsStoppingCriteria(stop_str, chatbot.tokenizer)
 
@@ -324,11 +327,10 @@ def generate_response(
 
     return {"status": "success", "response": text, "conversation_id": id(chatbot.conversation)}
 
-
 # ===================== Public API =====================
 
 def query(payload: dict):
-    """HF Endpoint entry (demo parity)."""
+    """HF Endpoint entry (demo parity + style hint)."""
     global model_initialized, tokenizer, model, image_processor, context_len, args
     if not model_initialized:
         if not initialize_model():
@@ -384,7 +386,6 @@ def get_model_info():
         "device": str(next(model.parameters()).device) if model else "Unknown",
     }
 
-
 # ===================== Init & Session =====================
 
 class _Args:
@@ -448,7 +449,6 @@ def initialize_model():
                 model_ = model_.to(torch.device("cuda"))
         model_.eval()
 
-        # assign globals
         globals()["tokenizer"] = tokenizer_
         globals()["model"] = model_
         globals()["image_processor"] = image_processor_
@@ -461,7 +461,6 @@ def initialize_model():
         print(f"[init] failed: {e}")
         return False
 
-
 # ===================== HF EndpointHandler =====================
 
 class EndpointHandler:
@@ -479,4 +478,4 @@ class EndpointHandler:
         return get_model_info()
 
 if __name__ == "__main__":
-    print("Handler ready (Demo Parity Mode). Use `EndpointHandler` or `query`.")
+    print("Handler ready (Demo Parity + Style Hint). Use `EndpointHandler` or `query`.")