no stop update

handler.py

@@ -1,11 +1,12 @@
 # -*- coding: utf-8 -*-
 """
-PULSE ECG Handler - Demo-like (sampling) + no_stop bayrağı
-- Demo davranışı: do_sample=True, temperature/top_p payload'dan
-- max_new_tokens: payload/slider değeri (KIRPMA YOK, direkt kullanılır)
-- İsteğe bağlı: no_stop=True ile stopping_criteria devre dışı
-- Tek görsel işleme; IM_START/END otomatik; 3D/4D/5D tensör uyumlu
-- Çıktıya post-format/deduplicate UYGULANMAZ (demo ile bire bir)
+PULSE ECG Handler - Demo-like sampling + no_stop (hard) + min_new_tokens auto
+- do_sample=True, temperature/top_p payload'dan
+- max_new_tokens: payload değeri (kırpma yok)
+- no_stop=True: stopping_criteria KAPALI + eos_token_id=None
+- no_stop=True ve min_new_tokens boşsa: otomatik min_new_tokens (uzun yanıt garantisi)
+- Tek görsel; IM_START/END otomatik; 3D/4D/5D tensör uyumlu
+- Post-format yok (demo davranışı)
 """
 
 import os
@@ -19,7 +20,7 @@ import torch
 from PIL import Image
 import requests
 
-# --- Opsiyonel bağımlılıklar ---
+# --- Opsiyonel ---
 try:
     import cv2
     CV2_AVAILABLE = True
@@ -27,7 +28,7 @@ except Exception:
     CV2_AVAILABLE = False
     print("Warning: OpenCV (cv2) not available; video is disabled.")
 
-# --- LLaVA / Transformers ---
+# --- LLaVA ---
 try:
     from llava.constants import (
         IMAGE_TOKEN_INDEX,
@@ -56,7 +57,6 @@ try:
 except Exception:
     HF_HUB_AVAILABLE = False
 
-# ------------- HF Hub init (opsiyonel) -------------
 api = None
 repo_name = ""
 if HF_HUB_AVAILABLE and "HF_TOKEN" in os.environ:
@@ -69,13 +69,11 @@ if HF_HUB_AVAILABLE and "HF_TOKEN" in os.environ:
         api = None
         repo_name = ""
 
-# ------------- Klasörler -------------
 LOGDIR = "./logs"
 VOTEDIR = "./votes"
 os.makedirs(LOGDIR, exist_ok=True)
 os.makedirs(VOTEDIR, exist_ok=True)
 
-# ------------- Global durum -------------
 tokenizer = None
 model = None
 image_processor = None
@@ -83,8 +81,6 @@ context_len = None
 args = None
 model_initialized = False
 
-# ------------- Yardımcılar -------------
-
 def _safe_upload(path: str):
     if api and repo_name and os.path.isfile(path):
         try:
@@ -100,17 +96,10 @@ def _safe_upload(path: str):
 def _conv_log_path():
     t = datetime.datetime.now()
     p = os.path.join(LOGDIR, f"{t.year:04d}-{t.month:02d}-{t.day:02d}-user_conv.json")
-    os.makedirs(os.path.dirname(p), exist_ok=True
-    )
+    os.makedirs(os.path.dirname(p), exist_ok=True)
     return p
 
 def load_image_any(image_input):
-    """
-    Desteklenen formatlar:
-      - URL (http/https)
-      - Yerel dosya yolu
-      - base64 (opsiyonel data URL prefix ile)
-    """
     if isinstance(image_input, str):
         s = image_input.strip()
         if s.startswith(("http://", "https://")):
@@ -119,14 +108,10 @@ def load_image_any(image_input):
             return Image.open(BytesIO(r.content)).convert("RGB")
         if os.path.exists(s):
             return Image.open(s).convert("RGB")
-        # base64
         if s.startswith("data:image"):
             s = s.split(",", 1)[1]
-        try:
-            raw = base64.b64decode(s)
-            return Image.open(BytesIO(raw)).convert("RGB")
-        except Exception as e:
-            raise ValueError(f"Invalid image string (not URL/path/base64): {e}")
+        raw = base64.b64decode(s)
+        return Image.open(BytesIO(raw)).convert("RGB")
     elif isinstance(image_input, dict) and "image" in image_input:
         return load_image_any(image_input["image"])
     else:
@@ -134,14 +119,10 @@ def load_image_any(image_input):
 
 def _guess_conv_mode(model_path: str) -> str:
     name = get_model_name_from_path(model_path).lower()
-    if "llama-2" in name:
-        return "llava_llama_2"
-    if "v1" in name or "pulse" in name:
-        return "llava_v1"
-    if "mpt" in name:
-        return "mpt"
-    if "qwen" in name:
-        return "qwen_1_5"
+    if "llama-2" in name: return "llava_llama_2"
+    if "v1" in name or "pulse" in name: return "llava_v1"
+    if "mpt" in name: return "mpt"
+    if "qwen" in name: return "qwen_1_5"
     return "llava_v0"
 
 def _wrap_image_token_if_needed(model_cfg) -> bool:
@@ -150,19 +131,15 @@ def _wrap_image_token_if_needed(model_cfg) -> bool:
     except Exception:
         return False
 
-# ------------- Çekirdek üretim -------------
-
 def _build_prompt_and_ids(chatbot, user_text: str, device: torch.device):
     use_wrap = _wrap_image_token_if_needed(chatbot.model.config)
     if use_wrap:
         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}"
-
     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)
@@ -184,54 +161,45 @@ def generate_response(
     conv_mode_override: str | None = None,
     det_seed: int | None = None,
     no_stop: bool = False,
-    min_new_tokens: int | None = None,   # opsiyonel, uzunluğu zorlamak istersen
+    min_new_tokens: int | None = None,   # otomatik atanabilir
 ):
     if not LLAVA_AVAILABLE:
         return {"error": "LLaVA modules not available"}
-
     if not message_text or image_input is None:
         return {"error": "Both 'message' and 'image' are required"}
 
-    # Chatbot/konuşma hazırla (her çağrıda sıfırdan, demo gibi)
     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()
     else:
         chatbot.conversation = conv_templates[chatbot.conv_mode].copy()
 
-    # Görüntüyü yükle
     try:
         pil_img = load_image_any(image_input)
     except Exception as e:
         return {"error": f"Failed to load image: {e}"}
 
-    # Log için kaydet
-    img_hash = "NA"
-    img_path = None
+    # log
+    img_hash, img_path = "NA", None
     try:
-        buf = BytesIO()
-        pil_img.save(buf, format="JPEG")
-        img_bytes = buf.getvalue()
+        buf = BytesIO(); pil_img.save(buf, format="JPEG"); img_bytes = buf.getvalue()
         img_hash = hashlib.md5(img_bytes).hexdigest()
         t = datetime.datetime.now()
         img_path = os.path.join(LOGDIR, "serve_images", f"{t.year:04d}-{t.month:02d}-{t.day:02d}", f"{img_hash}.jpg")
         os.makedirs(os.path.dirname(img_path), exist_ok=True)
-        if not os.path.isfile(img_path):
-            pil_img.save(img_path)
+        if not os.path.isfile(img_path): pil_img.save(img_path)
     except Exception as e:
         print(f"[log] saving image failed: {e}")
 
-    # Görüntüyü tensöre çevir
+    # görüntü tensörü
     device = next(chatbot.model.parameters()).device
     dtype  = next(chatbot.model.parameters()).dtype
     try:
         processed = process_images([pil_img], chatbot.image_processor, chatbot.model.config)
         if isinstance(processed, torch.Tensor):
-            if processed.ndim == 3:   # (C,H,W)
-                image_tensor = processed.unsqueeze(0)
-            elif processed.ndim == 4: # (B,C,H,W)
-                image_tensor = processed
-            elif processed.ndim == 5: # (B,T,C,H,W) -> (B*T,C,H,W)
+            if processed.ndim == 3:   image_tensor = processed.unsqueeze(0)
+            elif processed.ndim == 4: image_tensor = processed
+            elif processed.ndim == 5:
                 b,t,c,h,w = processed.shape
                 image_tensor = processed.reshape(b*t, c, h, w)
             else:
@@ -241,17 +209,15 @@ def generate_response(
             image_tensor = first.unsqueeze(0) if isinstance(first, torch.Tensor) and first.ndim == 3 else first
         else:
             return {"error": "Image processing returned empty"}
-
-        # Demo tarafında half + to(device) kalıbı yaygın
         image_tensor = image_tensor.to(device=device, dtype=dtype)
     except Exception as e:
         return {"error": f"Image processing failed: {e}"}
 
-    # Prompt & tokenizasyon
-    prompt, input_ids = _build_prompt_and_ids(chatbot, message_text, device)
+    # prompt & ids
+    _, input_ids = _build_prompt_and_ids(chatbot, message_text, device)
     stopping = None if no_stop else _stopping(chatbot, input_ids)
 
-    # (opsiyonel) deterministik sampling
+    # deterministik sample (opsiyonel)
     if det_seed is not None:
         try:
             det_seed = int(det_seed)
@@ -262,15 +228,15 @@ def generate_response(
         except Exception:
             pass
 
-    # EOS/PAD güvenli al
-    eos_id = chatbot.tokenizer.eos_token_id
-    if eos_id is None:
-        try:
-            eos_id = chatbot.tokenizer.convert_tokens_to_ids("</s>")
-        except Exception:
-            eos_id = 0
+    # EOS/PAD
+    eos_id = tokenizer.eos_token_id
+    pad_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else (eos_id if eos_id is not None else 0)
+    # no_stop=True ise eos'a göre durmayı tamamen kapat
+    if no_stop:
+        eos_for_gen = None
+    else:
+        eos_for_gen = eos_id
 
-    # generate kwargs (demo-like)
     gen_kwargs = dict(
         inputs=input_ids,
         images=image_tensor,
@@ -278,15 +244,25 @@ def generate_response(
         temperature=float(temperature),
         top_p=float(top_p),
         repetition_penalty=float(repetition_penalty),
-        max_new_tokens=int(max_new_tokens),  # KIRPMA YOK
+        max_new_tokens=int(max_new_tokens),
         use_cache=False,
-        pad_token_id=eos_id,
-        eos_token_id=eos_id,
+        pad_token_id=pad_id,
+        eos_token_id=eos_for_gen,
         length_penalty=1.0,
         early_stopping=False,
         stopping_criteria=None if no_stop else [stopping],
     )
-    if min_new_tokens is not None:
+
+    # min_new_tokens otomatik (no_stop=True ve kullanıcı vermediyse)
+    if no_stop and (min_new_tokens is None):
+        try:
+            req = int(max_new_tokens)
+            auto_min = max(300, min(req - 64, 1024))  # 300–1024 bandında güvenli
+            if auto_min > 0:
+                gen_kwargs["min_new_tokens"] = auto_min
+        except Exception:
+            pass
+    elif min_new_tokens is not None:
         try:
             mn = int(min_new_tokens)
             if mn > 0 and mn <= int(max_new_tokens):
@@ -294,19 +270,17 @@ def generate_response(
         except Exception:
             pass
 
-    # Üretim
+    # generate
     try:
         with torch.no_grad():
             outputs = chatbot.model.generate(**gen_kwargs)
         gen = outputs[0][input_ids.shape[1]:]
-        text = chatbot.tokenizer.decode(gen, skip_special_tokens=True)
-
-        # Konuşmaya yerleştir (demo gibi)
+        text = tokenizer.decode(gen, skip_special_tokens=True)
         chatbot.conversation.messages[-1][-1] = text
     except Exception as e:
         return {"error": f"Generation failed: {e}"}
 
-    # Log yaz
+    # log
     try:
         row = {
             "time": datetime.datetime.now().isoformat(),
@@ -318,20 +292,17 @@ def generate_response(
         }
         with open(_conv_log_path(), "a", encoding="utf-8") as f:
             f.write(json.dumps(row, ensure_ascii=False) + "\n")
-        _safe_upload(_conv_log_path())
-        if img_path:
-            _safe_upload(img_path)
+        _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)}
 
-# ------------- API Yüzeyi -------------
+# -------- API --------
 
 def query(payload: dict):
-    """HF Endpoint ana giriş noktası (demo uyumlu)"""
+    """HF Endpoint entry (demo-like)"""
     global model_initialized, tokenizer, model, image_processor, context_len, args
-
     if not model_initialized:
         if not initialize_model():
             return {"error": "Model initialization failed"}
@@ -340,37 +311,23 @@ def query(payload: dict):
     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'."}
 
-        if not message.strip():
-            return {"error": "Missing 'message' text"}
-        if image is None:
-            return {"error": "Missing 'image'. Use 'image', 'image_url', or 'img'."}
-
-        # Demo: slider benzeri parametreler
         max_new_tokens      = int(payload.get("max_output_tokens", payload.get("max_new_tokens", payload.get("max_tokens", 4096))))
         temperature         = float(payload.get("temperature", 0.05))
         top_p               = float(payload.get("top_p", 1.0))
         repetition_penalty  = float(payload.get("repetition_penalty", 1.0))
         conv_mode_override  = payload.get("conv_mode", None)
-
-        # (Opsiyonel) deterministik sample için seed
-        det_seed = payload.get("det_seed", 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
-
-        # (Yeni) stopping_criteria kapatma bayrağı
-        no_stop = bool(payload.get("no_stop", False))
-
-        # (Opsiyonel) min_new_tokens
-        mnt = payload.get("min_new_tokens", None)
+            try: det_seed = int(det_seed)
+            except Exception: det_seed = None
+        no_stop             = bool(payload.get("no_stop", False))
+        mnt                 = payload.get("min_new_tokens", None)
         if mnt is not None:
-            try:
-                mnt = int(mnt)
-            except Exception:
-                mnt = None
+            try: mnt = int(mnt)
+            except Exception: mnt = None
 
         return generate_response(
             message_text=message,
@@ -405,7 +362,7 @@ def get_model_info():
         "device": str(next(model.parameters()).device) if model else "Unknown",
     }
 
-# ------------- Model init -------------
+# -------- init --------
 
 class _Args:
     def __init__(self):
@@ -428,11 +385,8 @@ class InferenceDemo:
             tokenizer, model, image_processor, context_len
         )
         conv_mode_auto = _guess_conv_mode(model_path)
-        if args.conv_mode and args.conv_mode != conv_mode_auto:
-            self.conv_mode = args.conv_mode
-        else:
-            self.conv_mode = conv_mode_auto
-            args.conv_mode = conv_mode_auto
+        self.conv_mode = args.conv_mode if args.conv_mode else conv_mode_auto
+        args.conv_mode = self.conv_mode
         self.conversation = conv_templates[self.conv_mode].copy()
         self.num_frames = args.num_frames
 
@@ -453,7 +407,6 @@ class ChatSessionManager:
 chat_manager = ChatSessionManager()
 
 def initialize_model():
-    """Modeli yükle (lazy)"""
     global tokenizer, model, image_processor, context_len, args
     if not LLAVA_AVAILABLE:
         print("LLaVA not available; cannot init.")
@@ -464,14 +417,12 @@ 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
         )
-        # Cihaz
         try:
             _ = next(model.parameters()).device
         except Exception:
             if torch.cuda.is_available():
                 model = model.to(torch.device("cuda"))
         model.eval()
-        # Chatbot init
         chat_manager.init_if_needed(args, args.model_path, tokenizer, model, image_processor, context_len)
         print("[init] model/tokenizer/image_processor loaded.")
         return True
@@ -479,10 +430,7 @@ def initialize_model():
         print(f"[init] failed: {e}")
         return False
 
-# ------------- HF EndpointHandler -------------
-
 class EndpointHandler:
-    """Hugging Face Endpoint uyumlu sınıf"""
     def __init__(self, model_dir):
         self.model_dir = model_dir
         print(f"EndpointHandler initialized with model_dir: {model_dir}")