diff --git a/molt/utils/utils.py b/molt/utils/utils.py
index 76c3ff8..b523c3b 100644
--- a/molt/utils/utils.py
+++ b/molt/utils/utils.py
@@ -66,9 +66,12 @@ def get_tokenizer(pretrain, model, padding_side="left", use_fast=True):
         # AutoProcessor wraps tokenizer + image_processor; downstream code
         # detects VLM via hasattr(tokenizer, "image_processor").
         tokenizer = AutoProcessor.from_pretrained(pretrain, trust_remote_code=True)
+        # A checkpoint whose config declares vision_config but ships no complete
+        # processor bundle (e.g. no preprocessor_config.json) makes AutoProcessor
+        # fall back to returning the bare tokenizer; use it directly.
+        inner = getattr(tokenizer, "tokenizer", tokenizer)
         # AutoProcessor doesn't delegate tokenizer attributes, so set them on
         # the inner tokenizer and mirror the essentials back.
-        inner = tokenizer.tokenizer
         inner.padding_side = padding_side
         if inner.pad_token is None:
             inner.pad_token = inner.eos_token
diff --git a/tests/unit/test_get_tokenizer.py b/tests/unit/test_get_tokenizer.py
new file mode 100644
index 0000000..da0fcff
--- /dev/null
+++ b/tests/unit/test_get_tokenizer.py
@@ -0,0 +1,54 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+from types import SimpleNamespace
+
+from transformers import AutoProcessor
+
+from molt.utils.utils import get_tokenizer
+
+
+class _BareTokenizer:
+    """What AutoProcessor returns for a vision_config model without a full
+    processor bundle: the tokenizer itself, with no .tokenizer attribute."""
+
+    def __init__(self):
+        self.padding_side = "right"
+        self.pad_token = None
+        self.pad_token_id = None
+        self.eos_token = "</s>"
+        self.eos_token_id = 2
+
+
+class _Processor:
+    def __init__(self):
+        self.tokenizer = _BareTokenizer()
+
+
+def _get(monkeypatch, returned):
+    # Patch the classmethod rather than the module attribute: transformers'
+    # lazy module re-resolves attributes on import, which bypasses module-level
+    # monkeypatching.
+    monkeypatch.setattr(AutoProcessor, "from_pretrained", classmethod(lambda cls, *args, **kwargs: returned))
+    model = SimpleNamespace(is_vlm=True, config=SimpleNamespace(pad_token_id=None))
+    return get_tokenizer("some/vlm", model, padding_side="left")
+
+
+def test_vlm_processor_without_tokenizer_attribute_is_used_directly(monkeypatch):
+    bare = _BareTokenizer()
+    tokenizer = _get(monkeypatch, bare)
+
+    assert tokenizer is bare
+    assert tokenizer.padding_side == "left"
+    assert tokenizer.pad_token == "</s>"
+    assert tokenizer.pad_token_id == 2
+
+
+def test_vlm_processor_with_inner_tokenizer_mirrors_essentials(monkeypatch):
+    processor = _Processor()
+    tokenizer = _get(monkeypatch, processor)
+
+    assert tokenizer is processor
+    assert processor.tokenizer.padding_side == "left"
+    assert tokenizer.pad_token == "</s>"
+    assert tokenizer.pad_token_id == 2