Improve tokenizer loading fallback in bash/run.py

- Check whether the local tokenizer path actually exists before trying to load.
- Log clear warnings when local path is missing or fails.
- Respect USE_MODELSCOPE_HUB when falling back to remote model hub.
- Raise a descriptive RuntimeError instead of silently retrying HuggingFace
  and hanging on network-unreachable errors.

bash/run.py

@@ -40,6 +40,9 @@ import yaml
 from evalscope import run_task, TaskConfig
 from evalscope.api.agent import NativeAgentConfig
 from evalscope.config import SandboxTaskConfig
+from evalscope.utils.logger import get_logger
+
+logger = get_logger()
 
 SCRIPT_DIR = Path(__file__).parent.resolve()
 PROJECT_ROOT = SCRIPT_DIR.parent
@@ -281,13 +284,43 @@ _TOKENIZER = None
 
 def get_tokenizer(tokenizer_path: str):
     global _TOKENIZER
-    if _TOKENIZER is None:
+    if _TOKENIZER is not None:
+        return _TOKENIZER
+
     from transformers import AutoTokenizer
+    import os
+
+    # Try local path first
+    if tokenizer_path and os.path.exists(tokenizer_path):
         try:
             _TOKENIZER = AutoTokenizer.from_pretrained(tokenizer_path, trust_remote_code=True)
-        except Exception:
+            logger.info(f'Loaded tokenizer from local path: {tokenizer_path}')
-            _TOKENIZER = AutoTokenizer.from_pretrained('deepseek-ai/DeepSeek-V4-Flash', trust_remote_code=True)
             return _TOKENIZER
+        except Exception as e:
+            logger.warning(f'Failed to load tokenizer from local path {tokenizer_path}: {e}')
+
+    # If a local path was explicitly provided but does not exist, warn clearly
+    if tokenizer_path and tokenizer_path != '/data1/models/DeepSeek-V4-Flash-INT8':
+        logger.warning(f'Local tokenizer path does not exist: {tokenizer_path}. '
+                       f'Will try to download from model hub.')
+
+    # Fallback: try model hub. Respect ModelScope hub if configured.
+    fallback_model = 'deepseek-ai/DeepSeek-V4-Flash'
+    hub_source = 'ModelScope' if os.environ.get('USE_MODELSCOPE_HUB') == '1' else 'HuggingFace'
+    logger.info(f'Trying to load tokenizer from {hub_source}: {fallback_model}')
+    try:
+        _TOKENIZER = AutoTokenizer.from_pretrained(fallback_model, trust_remote_code=True)
+        logger.info(f'Loaded tokenizer from {hub_source}: {fallback_model}')
+        return _TOKENIZER
+    except Exception as e:
+        raise RuntimeError(
+            f'Failed to load tokenizer.\n'
+            f'  Local path: {tokenizer_path} (exists: {os.path.exists(tokenizer_path) if tokenizer_path else False})\n'
+            f'  {hub_source} fallback: {fallback_model}\n'
+            f'  Error: {e}\n'
+            f'  Hint: mount the tokenizer to the container and set --tokenizer-path correctly, '
+            f'or set USE_MODELSCOPE_HUB=1 and HF_ENDPOINT=https://hf-mirror.com if downloading.'
+        ) from e
 
 
 def truncate_middle(text: str, max_tokens: int, tokenizer_path: str) -> str: