Update app.py

app.py

@@ -1,72 +1,88 @@
 import os
+# Force CPU-only in this process by hiding CUDA devices (set before importing heavy libs)
+os.environ.setdefault('CUDA_VISIBLE_DEVICES', '')
+
 import torch
 import gradio as gr
 import time
-from ccd import ccd_eval, run_eval
-from libra.eval.run_libra import load_model
 
 # =========================================
 # Safe Libra Hook (CPU fallback + dtype fix)
+# This hook must run before any heavyweight libra model-loading occurs.
 # =========================================
-import torch
 import libra.model.builder as builder
 import libra.eval.run_libra as run_libra
 
-# 保存原始函数
-_original_load_pretrained_model = builder.load_pretrained_model
+# 保存原始函数（如果存在）
+_original_load_pretrained_model = getattr(builder, 'load_pretrained_model', None)
 
 def safe_load_pretrained_model(model_path, model_base=None, model_name=None, **kwargs):
     print("[INFO] Hook activated: safe_load_pretrained_model()")
 
-    # ---- 关键修复 1：补全 model_name，避免 .lower() on None ----
+    # 补全 model_name，避免 .lower() on None
     if model_name is None:
         model_name = model_path
 
-    # ---- 关键修复 2：强制以 CPU 参数调用原函数，彻底绕开 CUDA 初始化 ----
-    # 同时把 device_map 也设置为 cpu（避免传 'auto' 被塞进 {"": "auto"}）
-    kwargs = dict(kwargs)  # 避免原 dict 被上层复用
-    kwargs.setdefault("device", "cpu")
-    kwargs.setdefault("device_map", "cpu")
+    # 强制以 CPU 参数调用原函数，尽量避免 CUDA 初始化
+    kwargs = dict(kwargs)
+    kwargs.setdefault('device', 'cpu')
+    kwargs.setdefault('device_map', 'cpu')
 
-    # 注意：原函数内部仍会把 torch_dtype 设为 float16（除非 4/8bit），
-    # 但是我们可以在返回后统一上调为 float32。
-    tokenizer, model, image_processor, context_len = _original_load_pretrained_model(
-        model_path, model_base, model_name, **kwargs
-    )
+    if _original_load_pretrained_model is None:
+        raise RuntimeError('Original load_pretrained_model not found in builder')
 
-    # ---- 关键修复 3：CPU 环境统一上调到 float32，稳定运行 ----
+    # Try calling the original with our kwargs; if it doesn't accept them, fall back.
+    try:
+        tokenizer, model, image_processor, context_len = _original_load_pretrained_model(
+            model_path, model_base, model_name, **kwargs
+        )
+    except TypeError as te:
+        # Some implementations don't accept device/device_map kwargs. Retry without them.
+        print(f"[WARN] original load_pretrained_model rejected kwargs: {te} — retrying without device kwargs")
+        try:
+            tokenizer, model, image_processor, context_len = _original_load_pretrained_model(
+                model_path, model_base, model_name
+            )
+        except Exception as e:
+            print(f"[ERROR] load_pretrained_model failed on retry: {e}")
+            raise
+    except Exception:
+        # propagate other errors
+        raise
+
+    # 在 CPU 情况下尝试把模型和视觉塔上调到 float32，减少 CPU 上的兼容问题
     if not torch.cuda.is_available():
         try:
-            # 语言模型主体
             model.to(dtype=torch.float32)
         except Exception as e:
             print(f"[WARN] Could not upcast LM to float32: {e}")
-
         try:
-            # 视觉塔
             vt = model.get_vision_tower()
-            vt.to(device="cpu", dtype=torch.float32)
-            print("[INFO] Vision tower moved to cpu (float32).")
+            vt.to(device='cpu', dtype=torch.float32)
+            print('[INFO] Vision tower moved to cpu (float32).')
         except Exception as e:
             print(f"[WARN] Could not move vision_tower to cpu/float32: {e}")
     else:
-        # 若有 GPU，保持原来的 float16 + cuda（无需额外处理）
-        print("[INFO] GPU available — default CUDA fp16 path is kept.")
+        print('[INFO] GPU available — keeping original device/dtype behavior.')
 
     return tokenizer, model, image_processor, context_len
 
 # 将 builder 的加载函数替换为安全版
-builder.load_pretrained_model = safe_load_pretrained_model
+if _original_load_pretrained_model is not None:
+    builder.load_pretrained_model = safe_load_pretrained_model
 
-# 同时替换 run_libra.load_model，并把本地名也重绑定，确保后续调用走安全版
+# 同时替换 run_libra.load_model
 def safe_load_model(model_path, model_base=None, model_name=None):
-    print("[INFO] Hook activated: safe_load_model()")
+    print('[INFO] Hook activated: safe_load_model()')
     if model_name is None:
         model_name = model_path
     return safe_load_pretrained_model(model_path, model_base, model_name)
 
 run_libra.load_model = safe_load_model
-load_model = safe_load_model  # 让 app.py 后续的 load_model() 使用安全版
+
+# 现在导入 CCD 与其他被 hook 的符号（导入放在 hook 之后以确保生效）
+from ccd import ccd_eval, run_eval
+from libra.eval.run_libra import load_model
 
 # =========================================
 # Global Configuration
@@ -371,14 +387,16 @@ def main():
 
 
     # Log that Gradio is starting (helpful when stdout/stderr are captured)
+    # write startup log to local file in repository (avoid permission issues on Spaces)
     try:
-        with open('/workspace/CCD/callback.log', 'a', encoding='utf-8') as f:
+        os.makedirs('logs', exist_ok=True)
+        with open('logs/callback.log', 'a', encoding='utf-8') as f:
             f.write(f"\n=== GRADIO START ===\nstarted_at: {time.strftime('%Y-%m-%d %H:%M:%S')}\n\n")
     except Exception:
         pass
 
-    # Bind to 0.0.0.0 so the server is reachable from host/container and set an explicit port
-    demo.launch(share=True)
+
+    demo.launch()
 
 
 if __name__ == "__main__":