Update handler.py

handler.py

@@ -1,31 +1,106 @@
+import base64
+import io
+import os
+import requests
 import torch
 from PIL import Image
-from transformers import AutoProcessor, AutoModelForCausalLM
-import requests
+from transformers import pipeline
 
 class EndpointHandler:
-    def __init__(self, path=""):
-        model_id = "PULSE-ECG/PULSE-7B"
-        self.processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
-        self.model = AutoModelForCausalLM.from_pretrained(
-            model_id,
-            torch_dtype=torch.float16,
+    def __init__(self, path: str = ""):
+        # PULSE-7B (PyTorch/safetensors) + LLaVA tabanlı VLM
+        self.pipe = pipeline(
+            task="image-text-to-text",
+            model="PULSE-ECG/PULSE-7B",
             device_map="auto",
-            trust_remote_code=True
+            torch_dtype="auto",
+            trust_remote_code=True,
         )
+        # Uyarıları susturmak için pad_token
+        try:
+            if getattr(self.pipe.model.generation_config, "pad_token_id", None) is None:
+                self.pipe.model.generation_config.pad_token_id = self.pipe.model.config.eos_token_id
+        except Exception:
+            pass
+
+    # ---- yardımcılar ----
+    def _load_image(self, src: str) -> Image.Image:
+        """HTTP URL veya data URL (base64) kabul eder."""
+        if src.startswith("data:image/"):
+            # data:image/png;base64,xxxx
+            b64 = src.split(",", 1)[1]
+            return Image.open(io.BytesIO(base64.b64decode(b64))).convert("RGB")
+        # HTTP(S) URL
+        r = requests.get(src, stream=True, timeout=20)
+        r.raise_for_status()
+        return Image.open(r.raw).convert("RGB")
+
+    def _normalize_inputs(self, data: dict):
+        """
+        Aşağıdaki iki biçimi de destekler:
+        1) {"image_url": "...", "text": "..."}
+        2) {
+             "inputs": [
+               {"role":"user","content":[
+                 {"type":"image","image_url":"..."},
+                 {"type":"text","text":"..."}
+               ]}
+             ],
+             "parameters": {...}
+           }
+        """
+        # Biçim-1
+        if "image_url" in data or "text" in data:
+            image_url = data.get("image_url")
+            text = data.get("text", "Interpret this ECG image.")
+            if not image_url:
+                raise ValueError("No image_url provided")
+            return [
+                {"role": "user", "content": [
+                    {"type": "image", "image_url": image_url},
+                    {"type": "text", "text": text},
+                ]}
+            ], data.get("parameters", {})
+
+        # Biçim-2 (multimodal chat)
+        if "inputs" in data:
+            return data.get("inputs", []), data.get("parameters", {})
+
+        raise ValueError("Invalid payload: expected 'image_url'+'text' or 'inputs' format.")
+
+    # ---- giriş noktası ----
+    def __call__(self, data: dict):
+        """
+        Dönen sonuç, TGI/Toolkit ile uyumlu ham çıktı olur.
+        Client, aşağıdaki iki şemadan birini gönderebilir:
+
+        # Şema-1 (basit)
+        { "image_url": "https://.../ecg.png", "text": "Kısa yorum yaz", "parameters": {"max_new_tokens":256} }
 
-    def __call__(self, data: dict) -> dict:
-        image_url = data.get("image_url")
-        text = data.get("text", "Interpret this ECG image.")
+        # Şema-2 (multimodal chat)
+        {
+          "inputs": [
+            {"role":"user","content":[
+              {"type":"image","image_url":"https://.../ecg.png"},
+              {"type":"text","text":"Kısa yorum yaz"}
+            ]}
+          ],
+          "parameters": {"max_new_tokens":256, "temperature":0.2}
+        }
+        """
+        # 1) İstek şemasını normalize et
+        msgs, params = self._normalize_inputs(data)
 
-        if not image_url:
-            return {"error": "No image_url provided"}
+        # 2) Eğer image_url base64/URL ise pipe doğrudan kabul eder;
+        #    ekstra dönüştürme gerekmiyor (pipeline image-text-to-text bunu destekler).
+        #    Ancak bazı eski sürümlerde local image objesi istenirse yukarıdaki _load_image kullanılabilir.
 
-        image = Image.open(requests.get(image_url, stream=True).raw).convert("RGB")
-        inputs = self.processor(text=text, images=image, return_tensors="pt").to(self.model.device)
+        # 3) Varsayılan parametreler
+        params = {"max_new_tokens": 512, "temperature": 0.2, **(params or {})}
 
-        with torch.no_grad():
-            outputs = self.model.generate(**inputs, max_new_tokens=512)
+        # 4) Çalıştır
+        with torch.inference_mode():
+            out = self.pipe(msgs, **params)
 
-        result = self.processor.decode(outputs[0], skip_special_tokens=True)
-        return {"result": result}
+        # pipe, genellikle metni doğrudan string veya dict listesi olarak döndürür.
+        return out