update handler py

handler.py

@@ -1,5 +1,10 @@
-import base64, io, os, logging
-import requests, torch, transformers
+import base64
+import io
+import os
+import logging
+import requests
+import torch
+import transformers
 from PIL import Image
 from transformers import pipeline, AutoConfig, AutoModelForCausalLM, AutoProcessor
 
@@ -8,7 +13,7 @@ class EndpointHandler:
         logging.warning(f"[INIT] Transformers version: {transformers.__version__}")
         self.model_id = os.getenv("PULSE_MODEL_ID", "PULSE-ECG/PULSE-7B")
 
-        # 1) Normal yol: pipeline
+        # 1) Normal path: attempt pipeline directly
         try:
             self.pipe = pipeline(
                 task="image-text-to-text",
@@ -43,25 +48,33 @@ class EndpointHandler:
         except Exception as e:
             logging.warning(f"[INIT] override failed: {e}")
 
-        # 3) Fallback: AutoProcessor + AutoModel
-        logging.warning("[INIT] Fallback: AutoProcessor/AutoModel")
-        proc = AutoProcessor.from_pretrained(self.model_id, trust_remote_code=True)
-        mdl = AutoModelForCausalLM.from_pretrained(
-            self.model_id,
-            device_map="auto",
-            torch_dtype="auto",
-            trust_remote_code=True
-        )
+        # 3) Fallback: AutoProcessor + AutoModel with config override check
+        try:
+            cfg = AutoConfig.from_pretrained(self.model_id, trust_remote_code=True)
+            if getattr(cfg, "model_type", None) == "llava_llama":
+                logging.warning("[INIT] Fallback override: llava_llama -> llava")
+                cfg.model_type = "llava"
+            proc = AutoProcessor.from_pretrained(self.model_id, trust_remote_code=True, config=cfg)
+            mdl = AutoModelForCausalLM.from_pretrained(
+                self.model_id,
+                device_map="auto",
+                torch_dtype="auto",
+                trust_remote_code=True,
+                config=cfg,
+            )
 
-        def _mini_pipe(msgs, **params):
-            inputs = proc(msgs, return_tensors="pt").to(mdl.device)
-            gen_kwargs = {"max_new_tokens": 512, **params}
-            with torch.inference_mode():
-                out_ids = mdl.generate(**inputs, **gen_kwargs)
-            return proc.tokenizer.batch_decode(out_ids, skip_special_tokens=True)
+            def _mini_pipe(msgs, **params):
+                inputs = proc(msgs, return_tensors="pt").to(mdl.device)
+                gen_kwargs = {"max_new_tokens": 512, **params}
+                with torch.inference_mode():
+                    out_ids = mdl.generate(**inputs, **gen_kwargs)
+                return proc.tokenizer.batch_decode(out_ids, skip_special_tokens=True)
 
-        self.pipe = _mini_pipe
-        logging.warning("[INIT] Fallback loaded")
+            self.pipe = _mini_pipe
+            logging.warning("[INIT] Fallback loaded")
+        except Exception as e:
+            logging.error(f"[INIT] Fallback failed: {e}")
+            raise
 
     # ---- helpers ----
     def _ensure_pad_token(self):
@@ -74,7 +87,7 @@ class EndpointHandler:
             pass
 
     def _normalize_inputs(self, data: dict):
-        # Basit şema
+        # Simple schema
         if "image_url" in data or "text" in data:
             image_url = data.get("image_url")
             text = data.get("text", "Interpret this ECG image.")
@@ -87,7 +100,7 @@ class EndpointHandler:
                 ]}
             ], data.get("parameters", {})
 
-        # Multimodal chat şeması
+        # Multimodal chat schema
         if "inputs" in data:
             return data.get("inputs", []), data.get("parameters", {})