Update handler.py

handler.py

@@ -1,16 +1,14 @@
 # -*- coding: utf-8 -*-
 """
-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
+PULSE ECG Handler — Deterministik Versiyon
+- Üretim ayarları: do_sample=False (Tutarlı çıktı), temperature/top_p etkisiz
+- 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
-- Post-process: YALNIZCA whitespace/biçim normalizasyonu (yönetim/öneri cümleleri korunur)
+- Seed/deterministic KAPALI (do_sample=False ile determinizm sağlanır)
+- STYLE_HINT: demo üslubuna yaklaşmak için
+- Post-process: YALNIZCA whitespace/biçim normalizasyonu
 """
-
 import os
 import re
 import json
@@ -20,7 +18,6 @@ import datetime
 from io import BytesIO
 from threading import Thread
 from typing import Optional, Union
-
 import torch
 from PIL import Image
 import requests
@@ -129,10 +126,10 @@ def load_image_any(image_input: Union[str, dict]) -> Image.Image:
             s = s.split(",", 1)[1]
         raw = base64.b64decode(s)
         return Image.open(BytesIO(raw)).convert("RGB")
-
+    
     if isinstance(image_input, dict) and "image" in image_input:
         return load_image_any(image_input["image"])
-
+    
     raise ValueError("Unsupported image input format")
 
 def _normalize_whitespace(text: str) -> str:
@@ -161,7 +158,7 @@ class SafeKeywordsStoppingCriteria(StoppingCriteria):
         self.tokenizer = tokenizer
         tok = tokenizer(keyword, add_special_tokens=False, return_tensors="pt").input_ids[0]
         self.kw_ids = tok  # shape: (n,)
-
+    
     def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
         if input_ids is None or input_ids.shape[0] == 0:
             return False
@@ -193,11 +190,13 @@ class ChatSessionManager:
         self.chatbot = None
         self.args = None
         self.model_path = None
+        
     def init_if_needed(self, args, model_path, tokenizer, model, image_processor, context_len):
         if self.chatbot is None:
             self.args = args
             self.model_path = model_path
             self.chatbot = InferenceDemo(args, model_path, tokenizer, model, image_processor, context_len)
+            
     def get_chatbot(self, args, model_path, tokenizer, model, image_processor, context_len):
         self.init_if_needed(args, model_path, tokenizer, model, image_processor, context_len)
         # Her çağrıda taze template (demo gibi yeni tur)
@@ -212,7 +211,6 @@ def _build_prompt_and_ids(chatbot, user_text: str, device: torch.device):
     chatbot.conversation.append_message(chatbot.conversation.roles[0], inp)
     chatbot.conversation.append_message(chatbot.conversation.roles[1], None)
     prompt = chatbot.conversation.get_prompt()
-
     input_ids = tokenizer_image_token(
         prompt, chatbot.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt"
     ).unsqueeze(0).to(device)
@@ -222,35 +220,33 @@ def generate_response(
     message_text: str,
     image_input,
     *,
-    temperature: Optional[float] = None,
-    top_p: Optional[float] = None,
+    temperature: Optional[float] = None, # Deterministik modda yoksayılır
+    top_p: Optional[float] = None,       # Deterministik modda yoksayılır
     max_new_tokens: Optional[int] = None,
     conv_mode_override: Optional[str] = None,
     repetition_penalty: Optional[float] = None,
-    det_seed: Optional[int] = None,  # None → stokastik (demo gibi)
+    det_seed: Optional[int] = None,      # Deterministik modda yoksayılır
 ):
     if not (LLAVA_AVAILABLE and TRANSFORMERS_AVAILABLE):
         return {"error": "Required libraries not available (llava/transformers)"}
     if not message_text or image_input is None:
         return {"error": "Both 'message' and 'image' are required"}
-
-    # Varsayılanlar → demo
-    if temperature is None: temperature = 0.05
-    if top_p is None: top_p = 1.0
+    
+    # Varsayılanlar
     if max_new_tokens is None: max_new_tokens = 4096
     if repetition_penalty is None: repetition_penalty = 1.0  # etkisiz
-
+    
     # Chat session
     chatbot = chat_manager.get_chatbot(args, args.model_path, tokenizer, model, image_processor, context_len)
     if conv_mode_override and conv_mode_override in conv_templates:
         chatbot.conversation = conv_templates[conv_mode_override].copy()
-
+        
     # Görüntü yükle
     try:
         pil_img = load_image_any(image_input)
     except Exception as e:
         return {"error": f"Failed to load image: {e}"}
-
+        
     # Log için hash+path
     img_hash, img_path = "NA", None
     try:
@@ -263,11 +259,11 @@ def generate_response(
             pil_img.save(img_path)
     except Exception as e:
         print(f"[log] save image failed: {e}")
-
+        
     # Cihaz/dtype
     device = next(chatbot.model.parameters()).device
     dtype = torch.float16  # demo: half
-
+    
     # Görüntü ön-işleme → tensör
     try:
         processed = process_images([pil_img], chatbot.image_processor, chatbot.model.config)
@@ -277,22 +273,23 @@ def generate_response(
             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)
         image_tensor = image_tensor.to(device=device, dtype=dtype)  # demo: half + device
     except Exception as e:
         return {"error": f"Image processing failed: {e}"}
-
+        
     # STYLE_HINT ekle ve prompt hazırla
     msg = (message_text or "").strip()
     msg = f"{msg}\n\n{STYLE_HINT}"
     _, input_ids = _build_prompt_and_ids(chatbot, msg, device)
-
+    
     # 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)
-
-    # Seed (gönderilmediyse stokastik → demo gibi)
+    
+    # Seed (do_sample=False olduğu için önemsiz, ancak kodda bırakılabilir)
     if det_seed is not None:
         try:
             s = int(det_seed)
@@ -302,26 +299,31 @@ def generate_response(
                 torch.cuda.manual_seed_all(s)
         except Exception:
             pass
-
+            
     # Streamer (demo gibi)
     streamer = TextIteratorStreamer(
         chatbot.tokenizer, skip_prompt=True, skip_special_tokens=True
     )
-
-    # Generate kwargs — demo ayarları
+    
+    # Generate kwargs — Deterministik Ayarlar
     gen_kwargs = dict(
         inputs=input_ids,
         images=image_tensor,
         streamer=streamer,
-        do_sample=True,                     # DEMO
-        temperature=float(temperature),     # DEMO default 0.05
-        top_p=float(top_p),                 # DEMO default 1.0
-        max_new_tokens=int(max_new_tokens), # DEMO slider
-        repetition_penalty=float(repetition_penalty),  # default 1.0 → etkisiz
+        
+        # 🟢 ÖNEMLİ DEĞİŞİKLİK: Deterministiği (Tutarlılığı) Aç
+        do_sample=False,                     
+        
+        # temperature ve top_p ayarları artık yoksayılır
+        # temperature=float(temperature), 
+        # top_p=float(top_p),
+        
+        max_new_tokens=int(max_new_tokens), 
+        repetition_penalty=float(repetition_penalty), 
         use_cache=False,
-        stopping_criteria=[stopping],       # DEMO-benzeri durdurma
+        stopping_criteria=[stopping],
     )
-
+    
     # Üretim (arka thread) + akışı topla
     try:
         t = Thread(target=chatbot.model.generate, kwargs=gen_kwargs)
@@ -334,7 +336,7 @@ def generate_response(
         chatbot.conversation.messages[-1][-1] = text
     except Exception as e:
         return {"error": f"Generation failed: {e}"}
-
+        
     # Log
     try:
         row = {
@@ -350,7 +352,7 @@ def generate_response(
         _safe_upload(_conv_log_path()); _safe_upload(img_path or "")
     except Exception as e:
         print(f"[log] failed: {e}")
-
+        
     return {"status": "success", "response": text, "conversation_id": id(chatbot.conversation)}
 
 # ===================== Public API =====================
@@ -362,25 +364,24 @@ def query(payload: dict):
         if not initialize_model():
             return {"error": "Model initialization failed"}
         model_initialized = True
-
     try:
         message = payload.get("message") or payload.get("query") or payload.get("prompt") or payload.get("istem") or ""
         image   = payload.get("image") or payload.get("image_url") or payload.get("img") or None
+        
         if not message.strip(): return {"error": "Missing 'message' text"}
         if image is None:       return {"error": "Missing 'image'. Use 'image', 'image_url', or 'img'."}
-
-        # Demo varsayılanları — payload override edebilir
-        temperature        = float(payload.get("temperature", 0.05))
+        
+        # Deterministik modda temperature/top_p yoksayılır, ancak API uyumluluğu için tutulur
+        temperature        = float(payload.get("temperature", 0.0))  # Default 0.0
         top_p              = float(payload.get("top_p", 1.0))
         max_new_tokens     = int(payload.get("max_output_tokens", payload.get("max_new_tokens", payload.get("max_tokens", 4096))))
-        repetition_penalty = float(payload.get("repetition_penalty", 1.0))  # etkisiz default
-
+        repetition_penalty = float(payload.get("repetition_penalty", 1.0)) 
         conv_mode_override = payload.get("conv_mode", None)
         det_seed           = payload.get("det_seed", None)
         if det_seed is not None:
             try: det_seed = int(det_seed)
             except Exception: det_seed = None
-
+            
         return generate_response(
             message_text=message,
             image_input=image,
@@ -437,19 +438,18 @@ def initialize_model():
         tokenizer_, model_, image_processor_, context_len_ = load_pretrained_model(
             args.model_path, args.model_base, model_name, args.load_8bit, args.load_4bit
         )
-        # demo: model'ı genelde cuda’da çalıştırır
+        # model'ı cuda’ya taşı
         try:
             _ = next(model_.parameters()).device
         except Exception:
             if torch.cuda.is_available():
                 model_ = model_.to(torch.device("cuda"))
+        
         model_.eval()
-
         globals()["tokenizer"] = tokenizer_
         globals()["model"] = model_
         globals()["image_processor"] = image_processor_
         globals()["context_len"] = context_len_
-
         chat_manager.init_if_needed(args, args.model_path, tokenizer_, model_, image_processor_, context_len_)
         print("[init] model/tokenizer/image_processor loaded.")
         return True
@@ -464,14 +464,17 @@ class EndpointHandler:
     def __init__(self, model_dir):
         self.model_dir = model_dir
         print(f"EndpointHandler initialized with model_dir: {model_dir}")
+        
     def __call__(self, payload):
         if "inputs" in payload:
             return query(payload["inputs"])
         return query(payload)
+        
     def health_check(self):
         return health_check()
+        
     def get_model_info(self):
         return get_model_info()
 
 if __name__ == "__main__":
-    print("Handler ready (Demo Parity + Style Hint + whitespace post-process). Use `EndpointHandler` or `query`.")
+    print("Handler ready (Deterministik Mode: do_sample=False). Use `EndpointHandler` or `query`.")