Update handler.py

handler.py

@@ -1,10 +1,10 @@
 # -*- coding: utf-8 -*-
-# handler.py — Rapid_ECG / PULSE-7B — Lokal model yükleme + DEBUG
-# - YEREL model dizininden yükleme (HF Hub yok)
-# - HF Endpoint sözleşmesi (EndpointHandler(model_dir).load().__call__(inputs))
+# handler.py — Rapid_ECG / PULSE-7B — Stabil ve DEBUG'li sürüm (vision tower fix)
+# - HuggingFace Endpoint uyumlu
+# - Görsel sadece .preprocess() ile işlenir (process_images yok)
 # - {"inputs": {...}} sarmalaması destekli
-# - Sadece preprocess kullanır (process_images yok)
-# - Her kritik adımda [DEBUG] çıktı
+# - Vision tower kontrolü: mm_vision_tower veya vision_tower
+# - IMAGE_TOKEN_INDEX kullanımı
 
 import os
 import io
@@ -17,7 +17,7 @@ import torch
 from PIL import Image
 import requests
 
-# ===== LLaVA kurulumu (kütüphane) =====
+# ===== LLaVA kurulumu =====
 def _ensure_llava(tag: str = "v1.2.0"):
     try:
         import llava  # noqa
@@ -39,6 +39,7 @@ from llava.constants import (
     DEFAULT_IMAGE_TOKEN,
     DEFAULT_IM_START_TOKEN,
     DEFAULT_IM_END_TOKEN,
+    IMAGE_TOKEN_INDEX,   # <-- sabit index
 )
 from llava.model.builder import load_pretrained_model
 from llava.mm_utils import tokenizer_image_token
@@ -119,11 +120,6 @@ def _build_prompt_with_image(prompt: str, model_cfg) -> str:
         return f"{token}\n{prompt}"
     return f"{DEFAULT_IMAGE_TOKEN}\n{prompt}"
 
-def _require_files(dir_path: str, fnames: list):
-    missing = [f for f in fnames if not os.path.exists(os.path.join(dir_path, f))]
-    if missing:
-        raise FileNotFoundError(f"[ERROR] Missing files in {dir_path}: {missing}")
-
 # ---------- Endpoint Handler ----------
 class EndpointHandler:
     def __init__(self, model_dir: Optional[str] = None):
@@ -137,53 +133,38 @@ class EndpointHandler:
         self.dtype = _pick_dtype(self.device)
         self.model_name = None
 
-    def _resolve_local_model_dir(self) -> str:
-        # Öncelik: HF_MODEL_DIR env → self.model_dir → /repository
-        local = _get_env("HF_MODEL_DIR", None) or self.model_dir or "/repository"
-        local = os.path.abspath(local)
-        print(f"[DEBUG] resolved local model dir: {local}")
-        if not os.path.isdir(local):
-            raise FileNotFoundError(f"[ERROR] Local model directory not found: {local}")
-        return local
-
     def load(self):
-        # Yerel dizini çöz
-        model_path = self._resolve_local_model_dir()
-        # Bu dosyalar repo kökünde olmalı (sende var):
-        # - model.safetensors.index.json + shard'lar
-        # - tokenizer.model / tokenizer_config.json / config.json
-        _require_files(model_path, [
-            "config.json",
-            "tokenizer_config.json",
-            "tokenizer.model",
-            "model.safetensors.index.json",
-        ])
+        # Uzaktan yüklemek için HF_MODEL_ID; yerel için HF_MODEL_DIR kullanabilirsiniz (ayrı mantık eklemek isterseniz)
+        model_path = _get_env("HF_MODEL_ID", "PULSE-ECG/PULSE-7B")
+        model_base = _get_env("HF_MODEL_BASE", None)
+        print(f"[DEBUG] load(): HF_MODEL_ID={model_path}, HF_MODEL_BASE={model_base}")
 
         os.environ.setdefault("ATTN_IMPLEMENTATION", "flash_attention_2")
         os.environ.setdefault("FLASH_ATTENTION", "1")
 
-        print(f"[DEBUG] calling load_pretrained_model from local path: {model_path}")
-        try:
-            self.tokenizer, self.model, self.image_processor, self.context_len = load_pretrained_model(
-                model_path=model_path,
-                model_base=None,          # tam birleştirilmiş ağırlıklar için base yok
-                load_8bit=False,
-                load_4bit=False,
-                device_map="auto",
-                device=self.device,
-            )
-        except Exception as e:
-            raise RuntimeError(f"[ERROR] load_pretrained_model failed at {model_path}: {e}")
-
+        print("[DEBUG] calling load_pretrained_model ...")
+        self.tokenizer, self.model, self.image_processor, self.context_len = load_pretrained_model(
+            model_path=model_path,
+            model_base=model_base,
+            load_8bit=False,
+            load_4bit=False,
+            device_map="auto",
+            device=self.device,
+        )
         self.model_name = getattr(self.model.config, "name_or_path", str(model_path))
         print(f"[DEBUG] model loaded: name={self.model_name}")
 
-        # Vision tower kontrolü (yanlış model ise burada yakalar)
-        vt = getattr(self.model.config, "vision_tower", None)
+        # ---- Vision tower kontrolü: mm_vision_tower veya vision_tower
+        vt = (
+            getattr(self.model.config, "mm_vision_tower", None)
+            or getattr(self.model.config, "vision_tower", None)
+        )
+        print(f"[DEBUG] vision tower: {vt}")
         if self.image_processor is None or vt is None:
             raise RuntimeError(
-                f"[ERROR] Vision tower not loaded (vision_tower={vt}). "
-                f"Bu dizin multimodal (LLaVA/PULSE) bir model içermiyor gibi görünüyor: {model_path}"
+                "[ERROR] Vision tower not loaded (mm_vision_tower/vision_tower None). "
+                "Bu model multimodal değil veya yanlış checkpoint yüklendi. "
+                "HF_MODEL_ID olarak PULSE/LLaVA tabanlı bir model verin (örn: 'PULSE-ECG/PULSE-7B')."
             )
 
         # tokenizer güvenliği
@@ -242,10 +223,10 @@ class EndpointHandler:
         conv.append_message(conv.roles[1], None)
         full_prompt = conv.get_prompt()
 
-        # ---- tokenization
+        # ---- tokenization (IMAGE_TOKEN_INDEX ile)
         try:
             input_ids = tokenizer_image_token(
-                full_prompt, self.tokenizer, image_token_index=-200, return_tensors="pt"
+                full_prompt, self.tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors="pt"
             ).unsqueeze(0).to(self.device)
         except Exception:
             toks = self.tokenizer([full_prompt], return_tensors="pt", padding=True, truncation=True)
@@ -270,7 +251,7 @@ class EndpointHandler:
         except Exception as e:
             return {"error": f"Generation failed: {e}"}
 
-        # ---- decode
+        # ---- decode (sadece yeni tokenlar)
         new_tokens = gen_ids[0, input_ids.shape[1]:]
         text = self.tokenizer.decode(new_tokens, skip_special_tokens=True).strip()