Rename handler.bak.py to mm_utils_local.py

Browse files

Files changed (2) hide show

handler.bak.py +0 -540
mm_utils_local.py +259 -0

handler.bak.py DELETED Viewed

@@ -1,540 +0,0 @@
-"""
-PULSE-7B Enhanced Handler
-Ubden® Team - Edited by https://github.com/ck-cankurt
-Support: Text, Image URLs, and Base64 encoded images
-"""
-import torch
-from typing import Dict, List, Any
-import base64
-from io import BytesIO
-from PIL import Image
-import requests
-import time
-# Import utilities if available
-try:
-    from utils import (
-        performance_monitor,
-        validate_image_input,
-        sanitize_parameters,
-        get_system_info,
-        create_health_check,
-        deepseek_client
-    )
-    UTILS_AVAILABLE = True
-except ImportError:
-    UTILS_AVAILABLE = False
-    deepseek_client = None
-    print("⚠️ Utils module not found - performance monitoring and DeepSeek integration disabled")
-# Try to import LLaVA modules for proper conversation handling
-try:
-    from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN
-    from llava.conversation import conv_templates, SeparatorStyle
-    from llava.mm_utils import tokenizer_image_token, process_images, KeywordsStoppingCriteria
-    LLAVA_AVAILABLE = True
-    print("✅ LLaVA modules imported successfully")
-except ImportError:
-    LLAVA_AVAILABLE = False
-    print("⚠️ LLaVA modules not available - using basic text processing")
-class EndpointHandler:
-    def __init__(self, path=""):
-        """
-        Hey there! Let's get this PULSE-7B model up and running.
-        We'll load it from the HuggingFace hub directly, so no worries about local files.
-        Args:
-            path: Model directory path (we actually ignore this and load from HF hub)
-        """
-        print("🚀 Starting up PULSE-7B handler...")
-        print("📝 Enhanced by Ubden® Team - github.com/ck-cankurt")
-        import sys
-        print(f"🔧 Python version: {sys.version}")
-        print(f"🔧 PyTorch version: {torch.__version__}")
-        # Check transformers version
-        try:
-            import transformers
-            print(f"🔧 Transformers version: {transformers.__version__}")
-            # PULSE LLaVA works with transformers==4.37.2
-            if transformers.__version__ == "4.37.2":
-                print("✅ Using PULSE LLaVA compatible version (4.37.2)")
-            elif "dev" in transformers.__version__ or "git" in str(transformers.__version__):
-                print("⚠️ Using development version - may conflict with PULSE LLaVA")
-            else:
-                print("⚠️ Using different version - PULSE LLaVA prefers 4.37.2")
-        except Exception as e:
-            print(f"❌ Error checking transformers version: {e}")
-        print(f"🔧 CUDA available: {torch.cuda.is_available()}")
-        if torch.cuda.is_available():
-            print(f"🔧 CUDA device: {torch.cuda.get_device_name(0)}")
-        # Let's see what hardware we're working with
-        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-        print(f"🖥️ Running on: {self.device}")
-        try:
-            # First attempt - PULSE demo's exact approach
-            if LLAVA_AVAILABLE:
-                print("📦 Using PULSE demo's load_pretrained_model approach...")
-                from llava.model.builder import load_pretrained_model
-                from llava.mm_utils import get_model_name_from_path
-                model_path = "PULSE-ECG/PULSE-7B"
-                model_name = get_model_name_from_path(model_path)
-                self.tokenizer, self.model, self.image_processor, self.context_len = load_pretrained_model(
-                    model_path=model_path,
-                    model_base=None,
-                    model_name=model_name,
-                    load_8bit=False,
-                    load_4bit=False
-                )
-                # Move model to device like demo
-                self.model = self.model.to(self.device)
-                self.use_pipeline = False
-                print("✅ Model loaded successfully with PULSE demo's approach!")
-                print(f"📸 Image processor: {type(self.image_processor).__name__}")
-            else:
-                raise ImportError("LLaVA modules not available")
-        except Exception as e:
-            print(f"⚠️ PULSE demo approach failed: {e}")
-            print("🔄 Falling back to pipeline...")
-            try:
-                # Fallback - using pipeline
-                from transformers import pipeline
-                print("📦 Fetching model from HuggingFace Hub...")
-                self.pipe = pipeline(
-                    "text-generation",
-                    model="PULSE-ECG/PULSE-7B",
-                    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
-                    device=0 if torch.cuda.is_available() else -1,
-                    trust_remote_code=True,
-                    model_kwargs={
-                        "low_cpu_mem_usage": True,
-                        "use_safetensors": True
-                    }
-                )
-                self.use_pipeline = True
-                self.image_processor = None
-                print("✅ Model loaded successfully via pipeline!")
-            except Exception as e2:
-                print(f"😓 Pipeline also failed: {e2}")
-                try:
-                    # Last resort - manual loading
-                    from transformers import AutoTokenizer, LlamaForCausalLM
-                    print("📖 Manual loading as last resort...")
-                    self.tokenizer = AutoTokenizer.from_pretrained(
-                        "PULSE-ECG/PULSE-7B",
-                        trust_remote_code=True
-                    )
-                    self.model = LlamaForCausalLM.from_pretrained(
-                        "PULSE-ECG/PULSE-7B",
-                        torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
-                        device_map="auto",
-                        low_cpu_mem_usage=True,
-                        trust_remote_code=True
-                    )
-                    if self.tokenizer.pad_token is None:
-                        self.tokenizer.pad_token = self.tokenizer.eos_token
-                        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
-                    self.model.eval()
-                    self.use_pipeline = False
-                    self.image_processor = None
-                    print("✅ Model loaded manually!")
-                except Exception as e3:
-                    print(f"😓 All approaches failed: {e3}")
-                    self.pipe = None
-                    self.model = None
-                    self.tokenizer = None
-                    self.image_processor = None
-                    self.use_pipeline = None
-        # Final status report
-        print("\n🔍 Model Loading Status Report:")
-        print(f"   - use_pipeline: {self.use_pipeline}")
-        print(f"   - model: {'✅ Loaded' if hasattr(self, 'model') and self.model is not None else '❌ None'}")
-        print(f"   - tokenizer: {'✅ Loaded' if hasattr(self, 'tokenizer') and self.tokenizer is not None else '❌ None'}")
-        print(f"   - image_processor: {'✅ Loaded' if hasattr(self, 'image_processor') and self.image_processor is not None else '❌ None'}")
-        print(f"   - pipe: {'✅ Loaded' if hasattr(self, 'pipe') and self.pipe is not None else '❌ None'}")
-        # Check if any model component loaded successfully
-        has_model = hasattr(self, 'model') and self.model is not None
-        has_tokenizer = hasattr(self, 'tokenizer') and self.tokenizer is not None
-        has_pipe = hasattr(self, 'pipe') and self.pipe is not None
-        has_image_processor = hasattr(self, 'image_processor') and self.image_processor is not None
-        if not (has_model or has_tokenizer or has_pipe):
-            print("💥 CRITICAL: No model components loaded successfully!")
-        else:
-            print("✅ At least one model component loaded successfully")
-            if has_image_processor:
-                print("🖼️ Vision capabilities available!")
-            else:
-                print("⚠️ No image processor - text-only mode")
-    def is_valid_image_format(self, filename_or_url):
-        """Validate image format like PULSE demo"""
-        # Demo's supported formats
-        image_extensions = ["jpg", "jpeg", "png", "bmp", "gif", "tiff", "webp", "heic", "heif", "jfif", "svg", "eps", "raw"]
-        if filename_or_url.startswith(('http://', 'https://')):
-            # For URLs, check the extension or content-type
-            ext = filename_or_url.split('.')[-1].split('?')[0].lower()
-            return ext in image_extensions
-        else:
-            # For base64 or local files
-            return True  # Base64 will be validated during decode
-    def process_image_input(self, image_input):
-        """
-        Handle both URL and base64 image inputs exactly like PULSE demo
-        Args:
-            image_input: Can be a URL string or base64 encoded image
-        Returns:
-            PIL Image object or None if something goes wrong
-        """
-        try:
-            # Check if it's a URL (starts with http/https)
-            if isinstance(image_input, str) and (image_input.startswith('http://') or image_input.startswith('https://')):
-                print(f"🌐 Fetching image from URL: {image_input[:50]}...")
-                # Validate format like demo
-                if not self.is_valid_image_format(image_input):
-                    print("❌ Invalid image format in URL")
-                    return None
-                # Demo's exact image loading approach
-                response = requests.get(image_input, timeout=15)
-                if response.status_code == 200:
-                    image = Image.open(BytesIO(response.content)).convert("RGB")
-                    print(f"✅ Image downloaded successfully! Size: {image.size}")
-                    return image
-                else:
-                    print(f"❌ Failed to load image: status {response.status_code}")
-                    return None
-            # Must be base64 then
-            elif isinstance(image_input, str):
-                print("🔍 Decoding base64 image...")
-                # Remove the data URL prefix if it exists
-                base64_data = image_input
-                if "base64," in image_input:
-                    base64_data = image_input.split("base64,")[1]
-                # Clean and validate base64 data
-                base64_data = base64_data.strip().replace('\n', '').replace('\r', '').replace(' ', '')
-                try:
-                    image_data = base64.b64decode(base64_data)
-                    image = Image.open(BytesIO(image_data)).convert('RGB')
-                    print(f"✅ Base64 image decoded successfully! Size: {image.size}")
-                    return image
-                except Exception as decode_error:
-                    print(f"❌ Base64 decode error: {decode_error}")
-                    return None
-        except Exception as e:
-            print(f"❌ Couldn't process the image: {e}")
-            return None
-        return None
-    def add_turkish_commentary(self, response: Dict[str, Any], enable_commentary: bool, timeout: int = 30) -> Dict[str, Any]:
-        """Add Turkish commentary to the response using DeepSeek API"""
-        if not enable_commentary:
-            return response
-        if not UTILS_AVAILABLE or not deepseek_client:
-            print("⚠️ DeepSeek client not available - skipping Turkish commentary")
-            response["commentary_status"] = "unavailable"
-            return response
-        if not deepseek_client.is_available():
-            print("⚠️ DeepSeek API key not configured - skipping Turkish commentary")
-            response["commentary_status"] = "api_key_missing"
-            return response
-        generated_text = response.get("generated_text", "")
-        if not generated_text:
-            print("⚠️ No generated text to comment on")
-            response["commentary_status"] = "no_text"
-            return response
-        print("🔄 DeepSeek ile Türkçe yorum ekleniyor...")
-        commentary_result = deepseek_client.get_turkish_commentary(generated_text, timeout)
-        if commentary_result["success"]:
-            response["comment_text"] = commentary_result["comment_text"]
-            response["commentary_model"] = commentary_result.get("model", "deepseek-chat")
-            response["commentary_tokens"] = commentary_result.get("tokens_used", 0)
-            response["commentary_status"] = "success"
-            print("✅ Türkçe yorum başarıyla eklendi")
-        else:
-            response["comment_text"] = ""
-            response["commentary_error"] = commentary_result["error"]
-            response["commentary_status"] = "failed"
-            print(f"❌ Türkçe yorum eklenemedi: {commentary_result['error']}")
-        return response
-    def health_check(self) -> Dict[str, Any]:
-        """Health check endpoint"""
-        if UTILS_AVAILABLE:
-            return create_health_check()
-        else:
-            return {
-                'status': 'healthy',
-                'model': 'PULSE-7B',
-                'timestamp': time.time(),
-                'handler_version': '2.0.0'
-            }
-    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
-        """
-        Main processing function - where the magic happens!
-        Args:
-            data: Input data with 'inputs' and optional 'parameters'
-        Returns:
-            List with the generated response
-        """
-        # Quick check - is our model ready?
-        if self.use_pipeline is None:
-            return [{
-                "generated_text": "Oops! Model couldn't load properly. Please check the deployment settings.",
-                "error": "Model initialization failed",
-                "handler": "Ubden® Team Enhanced Handler"
-            }]
-        try:
-            # Parse the inputs - flexible format support
-            inputs = data.get("inputs", "")
-            text = ""
-            image = None
-            if isinstance(inputs, dict):
-                # Dictionary input - check for text and image
-                # Support query field (new) plus original text/prompt fields
-                text = inputs.get("query", inputs.get("text", inputs.get("prompt", str(inputs))))
-                # Check for image in various formats
-                image_input = inputs.get("image", inputs.get("image_url", inputs.get("image_base64", None)))
-                if image_input:
-                    image = self.process_image_input(image_input)
-                    if image:
-                        # Since we're in text-only mode, create smart ECG context
-                        print(f"🖼️ Image loaded: {image.size[0]}x{image.size[1]} pixels - using text-only ECG analysis mode")
-                        # Create ECG-specific prompt that mimics visual analysis
-                        ecg_context = f"Analyzing an ECG image ({image.size[0]}x{image.size[1]} pixels). "
-                        # Use demo's exact approach - no additional context, just the query
-                        # Model is trained to understand ECG images from text queries
-                        pass  # Keep text exactly as received
-            else:
-                # Simple string input
-                text = str(inputs)
-            if not text:
-                return [{"generated_text": "Hey, I need some text to work with! Please provide an input."}]
-            # Get generation parameters - using PULSE-7B demo's exact settings
-            parameters = data.get("parameters", {})
-            max_new_tokens = min(parameters.get("max_new_tokens", 1024), 8192)  # Demo uses 1024 default
-            temperature = parameters.get("temperature", 0.05)  # Demo uses 0.05 for precise medical analysis
-            top_p = parameters.get("top_p", 1.0)  # Demo uses 1.0 for full vocabulary access
-            do_sample = parameters.get("do_sample", True)  # Demo uses sampling
-            repetition_penalty = parameters.get("repetition_penalty", 1.0)  # Demo default
-            print(f"🎛️ Generation params: max_tokens={max_new_tokens}, temp={temperature}, top_p={top_p}, do_sample={do_sample}, rep_penalty={repetition_penalty}")
-            # Check if Turkish commentary is requested (NEW FEATURE)
-            enable_turkish_commentary = parameters.get("enable_turkish_commentary", False)  # Default false
-            # Using pipeline? Let's go!
-            if self.use_pipeline:
-                print(f"🎛️ Pipeline generation: temp={temperature}, tokens={max_new_tokens}")
-                print(f"📝 Input text: '{text[:100]}...'")
-                result = self.pipe(
-                    text,
-                    max_new_tokens=max_new_tokens,
-                                            min_new_tokens=200,  # Force very detailed analysis to match demo
-                    temperature=temperature,
-                    top_p=top_p,
-                    do_sample=do_sample,
-                    repetition_penalty=repetition_penalty,
-                    return_full_text=False  # Just the new stuff, not the input
-                )
-                # Pipeline returns a list, let's handle it
-                if isinstance(result, list) and len(result) > 0:
-                    generated_text = result[0].get("generated_text", "").strip()
-                    print(f"🔍 Pipeline debug:")
-                    print(f"   - Raw result: '{str(result[0])[:200]}...'")
-                    print(f"   - Generated text length: {len(generated_text)}")
-                    # Clean up common issues
-                    if generated_text.startswith(text):
-                        generated_text = generated_text[len(text):].strip()
-                        print("🔧 Removed input text from output")
-                    # Remove common artifacts
-                    generated_text = generated_text.replace("</s>", "").strip()
-                    if not generated_text:
-                        print("❌ Pipeline generated empty text!")
-                        generated_text = "Empty response from pipeline. Please try different parameters."
-                    print(f"✅ Final pipeline text: '{generated_text[:100]}...' (length: {len(generated_text)})")
-                    # Create response
-                    response = {"generated_text": generated_text}
-                    # Add Turkish commentary if requested (NEW FEATURE)
-                    if enable_turkish_commentary:
-                        response = self.add_turkish_commentary(response, True)
-                    return [response]
-                else:
-                    generated_text = str(result).strip()
-                    # Create response
-                    response = {"generated_text": generated_text}
-                    # Add Turkish commentary if requested (NEW FEATURE)
-                    if enable_turkish_commentary:
-                        response = self.add_turkish_commentary(response, True)
-                    return [response]
-            # Manual generation mode - using PULSE demo's exact approach
-            else:
-                print(f"🔥 Manual generation with PULSE demo logic: temp={temperature}, tokens={max_new_tokens}")
-                print(f"📝 Input text: '{text[:100]}...'")
-                # Text-only generation with enhanced ECG context
-                print("🔤 Using enhanced text-only generation with ECG context")
-                # Tokenize the enhanced prompt
-                encoded = self.tokenizer(
-                    text,
-                    return_tensors="pt",
-                    truncation=True,
-                    max_length=4096  # Increased for longer prompts
-                )
-                input_ids = encoded["input_ids"].to(self.device)
-                attention_mask = encoded.get("attention_mask")
-                if attention_mask is not None:
-                    attention_mask = attention_mask.to(self.device)
-                print(f"🔍 Enhanced generation debug:")
-                print(f"   - Enhanced prompt length: {len(text)} chars")
-                print(f"   - Input tokens: {input_ids.shape[-1]}")
-                print(f"   - Prompt preview: '{text[:150]}...'")
-                # Generate with enhanced settings for medical analysis
-                with torch.no_grad():
-                    outputs = self.model.generate(
-                        input_ids,
-                        attention_mask=attention_mask,
-                        max_new_tokens=max_new_tokens,
-                        min_new_tokens=200,  # Force detailed response like demo
-                        temperature=temperature,
-                        top_p=top_p,
-                        do_sample=do_sample,
-                        repetition_penalty=repetition_penalty,
-                        pad_token_id=self.tokenizer.pad_token_id,
-                        eos_token_id=self.tokenizer.eos_token_id,
-                        early_stopping=False
-                    )
-                # Decode and clean response
-                generated_ids = outputs[0][input_ids.shape[-1]:]
-                generated_text = self.tokenizer.decode(
-                    generated_ids,
-                    skip_special_tokens=True,
-                    clean_up_tokenization_spaces=True
-                ).strip()
-                # Aggressive cleanup of artifacts
-                generated_text = generated_text.replace("</s>", "").strip()
-                # Simple cleanup - just remove Answer prefix and parentheses
-                if generated_text.startswith("(Answer:") and ")" in generated_text:
-                    # Just remove the parentheses and Answer: prefix
-                    end_paren = generated_text.find(")")
-                    answer_content = generated_text[8:end_paren].strip()  # Remove "(Answer:"
-                    # Keep the rest of the response if there is any
-                    rest_of_response = generated_text[end_paren+1:].strip()
-                    if rest_of_response:
-                        generated_text = f"{answer_content}. {rest_of_response}"
-                    else:
-                        generated_text = answer_content
-                elif generated_text.startswith("Answer:"):
-                    generated_text = generated_text[7:].strip()
-                # Remove only clear training artifacts
-                cleanup_patterns = [
-                    "In this task",
-                    "I'm asking the respondent",
-                    "The respondent should"
-                ]
-                for pattern in cleanup_patterns:
-                    if pattern in generated_text:
-                        parts = generated_text.split(pattern)
-                        generated_text = parts[0].strip()
-                        break
-                # Only provide fallback if response is truly empty or malformed
-                if len(generated_text) < 10 or generated_text.startswith("7)"):
-                    print("⚠️ Malformed response detected, providing fallback...")
-                    generated_text = "This ECG shows cardiac electrical activity. For accurate interpretation, please consult with a qualified cardiologist who can analyze the specific waveforms, intervals, and morphology patterns."
-                print(f"✅ Enhanced text-only generation: '{generated_text[:100]}...' (length: {len(generated_text)})")
-                # Create response
-                response = {"generated_text": generated_text}
-                # Add Turkish commentary if requested (NEW FEATURE)
-                if enable_turkish_commentary:
-                    response = self.add_turkish_commentary(response, True)
-                return [response]
-        except Exception as e:
-            error_msg = f"Something went wrong during generation: {str(e)}"
-            print(f"❌ {error_msg}")
-            return [{
-                "generated_text": "",
-                "error": error_msg,
-                "handler": "Ubden® Team Enhanced Handler"
-            }]

mm_utils_local.py ADDED Viewed

	@@ -0,0 +1,259 @@

+# mm_utils_local.py
+# LLaVA/PULSE uyumlu, dayanıklı mm_utils (anyres + pad)
+# - crop_size/size alanlarını güvenli okur
+# - preprocess veya __call__ farkını soğurur
+# - patch_size'a tam bölünecek pad ekler
+# - upstream imzalarıyla uyumludur
+from typing import Any, Dict, List, Optional, Sequence, Tuple
+from io import BytesIO
+import base64
+import math
+import ast
+import torch
+from PIL import Image
+from transformers import StoppingCriteria
+from llava.constants import IMAGE_TOKEN_INDEX  # imza uyumu için
+# ---------- Yardımcılar ----------
+def _get_crop_size(processor: Any, default: int = 224) -> int:
+    cs = getattr(processor, "crop_size", None)
+    if cs is None:
+        sz = getattr(processor, "size", None)
+        if isinstance(sz, dict):
+            return int(sz.get("shortest_edge", default))
+        if isinstance(sz, int):
+            return int(sz)
+        return int(default)
+    if isinstance(cs, dict):
+        if "height" in cs:
+            return int(cs["height"])
+        if "shortest_edge" in cs:
+            return int(cs["shortest_edge"])
+        # beklenmedik dict: ilk değeri al
+        for v in cs.values():
+            return int(v)
+    return int(cs)
+def _get_shortest_edge(processor: Any, fallback: Optional[int] = None) -> int:
+    sz = getattr(processor, "size", None)
+    if isinstance(sz, dict) and "shortest_edge" in sz:
+        return int(sz["shortest_edge"])
+    if isinstance(sz, int):
+        return int(sz)
+    return _get_crop_size(processor, default=(fallback or 224))
+def _preprocess_one(processor: Any, img: Image.Image) -> torch.Tensor:
+    # Bazı sürümlerde .preprocess yok; direkt __call__ çalıştırılır.
+    if hasattr(processor, "preprocess"):
+        out = processor.preprocess(img, return_tensors="pt")
+    else:
+        out = processor(img, return_tensors="pt")
+    return out["pixel_values"][0]
+def pad_to_multiple(image: Image.Image, multiple: int) -> Image.Image:
+    w, h = image.size
+    W = math.ceil(w / multiple) * multiple
+    H = math.ceil(h / multiple) * multiple
+    if (W, H) == (w, h):
+        return image
+    canvas = Image.new(image.mode, (W, H), (0, 0, 0))
+    canvas.paste(image, (0, 0))
+    return canvas
+# ---------- Orijinal API ----------
+def select_best_resolution(original_size: Tuple[int, int], possible_resolutions: List[Tuple[int, int]]) -> Tuple[int, int]:
+    """Upstream ile aynı mantık: en etkili ve en az boşa giden çözünürlüğü seç."""
+    original_width, original_height = original_size
+    best_fit = None
+    max_effective_resolution = 0
+    min_wasted_resolution = float("inf")
+    for width, height in possible_resolutions:
+        scale = min(width / original_width, height / original_height)
+        down_w, down_h = int(original_width * scale), int(original_height * scale)
+        effective = min(down_w * down_h, original_width * original_height)
+        wasted = (width * height) - effective
+        if (effective > max_effective_resolution) or (effective == max_effective_resolution and wasted < min_wasted_resolution):
+            max_effective_resolution = effective
+            min_wasted_resolution = wasted
+            best_fit = (width, height)
+    return best_fit
+def resize_and_pad_image(image: Image.Image, target_resolution: Tuple[int, int]) -> Image.Image:
+    """Hedef çözünürlüğe orantıyı koruyarak resize + siyah pad."""
+    ow, oh = image.size
+    W, H = target_resolution
+    sw, sh = W / ow, H / oh
+    if sw < sh:
+        nw, nh = W, min(math.ceil(oh * sw), H)
+    else:
+        nh, nw = H, min(math.ceil(ow * sh), W)
+    resized = image.resize((nw, nh))
+    canvas = Image.new("RGB", (W, H), (0, 0, 0))
+    canvas.paste(resized, ((W - nw) // 2, (H - nh) // 2))
+    return canvas
+def divide_to_patches(image: Image.Image, patch_size: int) -> List[Image.Image]:
+    """Görüntüyü patch_size x patch_size karelere böl."""
+    patches: List[Image.Image] = []
+    W, H = image.size
+    for y in range(0, H, patch_size):
+        for x in range(0, W, patch_size):
+            patches.append(image.crop((x, y, x + patch_size, y + patch_size)))
+    return patches
+def get_anyres_image_grid_shape(image_size: Tuple[int, int], grid_pinpoints, patch_size: int) -> Tuple[int, int]:
+    """AnyRes sonrası patch ızgara boyutu (W//patch, H//patch)."""
+    if isinstance(grid_pinpoints, list):
+        possible_resolutions = grid_pinpoints
+    else:
+        possible_resolutions = ast.literal_eval(grid_pinpoints)
+    width, height = select_best_resolution(image_size, possible_resolutions)
+    return width // patch_size, height // patch_size
+def process_anyres_image(image: Image.Image, processor: Any, grid_pinpoints) -> torch.Tensor:
+    """
+    Robust AnyRes:
+      - crop_size/size güvenli okuma
+      - hedef çözünürlüğe resize+pad
+      - patch_size'a tam bölünecek pad
+      - preprocess/call farkını soyutlama
+    """
+    if isinstance(grid_pinpoints, list):
+        possible_resolutions = grid_pinpoints
+    else:
+        possible_resolutions = ast.literal_eval(grid_pinpoints)
+    patch_size = _get_crop_size(processor, default=224)
+    shortest_edge = _get_shortest_edge(processor, fallback=patch_size)
+    best_resolution = select_best_resolution(image.size, possible_resolutions)
+    image_padded = resize_and_pad_image(image, best_resolution)
+    image_padded = pad_to_multiple(image_padded, patch_size)
+    patches = divide_to_patches(image_padded, patch_size)
+    image_original_resize = image.resize((shortest_edge, shortest_edge))
+    image_patches = [_preprocess_one(processor, image_original_resize)]
+    image_patches += [_preprocess_one(processor, p) for p in patches]
+    return torch.stack(image_patches, dim=0)
+def load_image_from_base64(image: str) -> Image.Image:
+    return Image.open(BytesIO(base64.b64decode(image)))
+def expand2square(pil_img: Image.Image, background_color: Tuple[int, int, int]) -> Image.Image:
+    w, h = pil_img.size
+    if w == h:
+        return pil_img
+    if w > h:
+        result = Image.new(pil_img.mode, (w, w), background_color)
+        result.paste(pil_img, (0, (w - h) // 2))
+        return result
+    result = Image.new(pil_img.mode, (h, h), background_color)
+    result.paste(pil_img, ((h - w) // 2, 0))
+    return result
+def process_images(images: List[Image.Image], image_processor: Any, model_cfg: Any):
+    """
+    Upstream API ile aynı isim/geri dönüş; ancak daha dayanıklı:
+      - pad: image_mean yoksa güvenli varsayılan (0.5,0.5,0.5)
+      - anyres: robust process_anyres_image
+      - else: toplu çağrı TypeError ise tek tek çağrı fallback
+    """
+    # bazı konfig’lerde alan adı mm_image_aspect_ratio olabilir
+    image_aspect_ratio = getattr(model_cfg, "image_aspect_ratio", None) or getattr(model_cfg, "mm_image_aspect_ratio", None)
+    new_images: List[torch.Tensor] = []
+    if image_aspect_ratio == "pad":
+        for image in images:
+            img_mean = getattr(image_processor, "image_mean", [0.5, 0.5, 0.5])
+            bg = tuple(int(x * 255) for x in img_mean)
+            image_sq = expand2square(image, bg)
+            image_t = _preprocess_one(image_processor, image_sq)
+            new_images.append(image_t)
+    elif image_aspect_ratio == "anyres":
+        grid = getattr(model_cfg, "image_grid_pinpoints", "[(336,336)]")
+        for image in images:
+            image_t = process_anyres_image(image, image_processor, grid)
+            new_images.append(image_t)
+    else:
+        try:
+            out = image_processor(images, return_tensors="pt")
+            return out["pixel_values"]
+        except TypeError:
+            outs = [image_processor(img, return_tensors="pt") for img in images]
+            pix = [o["pixel_values"][0] for o in outs]
+            return torch.stack(pix, dim=0)
+    if all(x.shape == new_images[0].shape for x in new_images):
+        return torch.stack(new_images, dim=0)
+    return new_images
+def tokenizer_image_token(prompt: str, tokenizer: Any, image_token_index: int = IMAGE_TOKEN_INDEX, return_tensors: Optional[str] = None):
+    prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split("<image>")]
+    def insert_separator(X, sep):
+        return [ele for sublist in zip(X, [sep] * len(X)) for ele in sublist][:-1]
+    input_ids: List[int] = []
+    offset = 0
+    if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
+        offset = 1
+        input_ids.append(prompt_chunks[0][0])
+    for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
+        input_ids.extend(x[offset:])
+    if return_tensors is not None:
+        if return_tensors == "pt":
+            return torch.tensor(input_ids, dtype=torch.long)
+        raise ValueError(f"Unsupported tensor type: {return_tensors}")
+    return input_ids
+def get_model_name_from_path(model_path: str) -> str:
+    model_path = model_path.strip("/")
+    model_paths = model_path.split("/")
+    if model_paths[-1].startswith("checkpoint-"):
+        return model_paths[-2] + "_" + model_paths[-1]
+    else:
+        return model_paths[-1]
+# Upstream ile uyumlu: durdurma kriteri
+class KeywordsStoppingCriteria(StoppingCriteria):
+    def __init__(self, keywords, tokenizer, input_ids):
+        self.keywords = keywords
+        self.keyword_ids = []
+        self.max_keyword_len = 0
+        for keyword in keywords:
+            cur_keyword_ids = tokenizer(keyword).input_ids
+            if len(cur_keyword_ids) > 1 and cur_keyword_ids[0] == tokenizer.bos_token_id:
+                cur_keyword_ids = cur_keyword_ids[1:]
+            if len(cur_keyword_ids) > self.max_keyword_len:
+                self.max_keyword_len = len(cur_keyword_ids)
+            self.keyword_ids.append(torch.tensor(cur_keyword_ids))
+        self.tokenizer = tokenizer
+        self.start_len = input_ids.shape[1]
+    def call_for_batch(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        offset = min(output_ids.shape[1] - self.start_len, self.max_keyword_len)
+        self.keyword_ids = [kid.to(output_ids.device) for kid in self.keyword_ids]
+        for kid in self.keyword_ids:
+            truncated = output_ids[0, -kid.shape[0]:]
+            if torch.equal(truncated, kid):
+                return True
+        outputs = self.tokenizer.batch_decode(output_ids[:, -offset:], skip_special_tokens=True)[0]
+        for keyword in self.keywords:
+            if keyword in outputs:
+                return True
+        return False
+    def __call__(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        outs = []
+        for i in range(output_ids.shape[0]):
+            outs.append(self.call_for_batch(output_ids[i].unsqueeze(0), scores))
+        return all(outs)