From b4734b9550eb85cb8e38e3d93ef8fccb83ddf8fe Mon Sep 17 00:00:00 2001
From: Alex Brooks <alex.brooks@ibm.com>
Date: Thu, 20 Nov 2025 22:32:30 -0700
Subject: [PATCH] [Bugfix] Fix default MM LoRA alignment for single str prompts
 (#29140)

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>
---
 tests/lora/test_default_mm_loras.py | 35 +++++++++++++++++++++++++++++
 vllm/entrypoints/llm.py             |  2 +-
 2 files changed, 36 insertions(+), 1 deletion(-)

diff --git a/tests/lora/test_default_mm_loras.py b/tests/lora/test_default_mm_loras.py
index dfc45e78e464f..407b29fdd1d58 100644
--- a/tests/lora/test_default_mm_loras.py
+++ b/tests/lora/test_default_mm_loras.py
@@ -5,7 +5,9 @@ Tests for applying default registered multimodal loras.
 """
 
 import os
+import unittest.mock as mock
 
+import pytest
 from huggingface_hub import snapshot_download
 
 from vllm.lora.request import LoRARequest
@@ -114,3 +116,36 @@ def test_default_mm_lora_fails_with_overridden_lora_request(
         default_mm_loras={"audio": IMAGE_LORA_PATH},
         expected_suffix=RESPONSE_SUFFIX_WITH_LORA,
     )
+
+
+def test_default_mm_lora_does_not_expand_string_reqs(vllm_runner):
+    class MockEngineException(Exception):
+        pass
+
+    # Regression test for ensuring default multimodal lora resolution
+    # does not expand the lora req if the prompt type is a string.
+    vllm_runner_kwargs = {
+        **VLLM_RUNNER_BASE_KWARGS,
+        **{"default_mm_loras": {"audio": AUDIO_LORA_PATH}},
+    }
+
+    # Avoid the full generation call since these tests are expensive;
+    # just check what lora request is actually submitted to the engine
+    mock_err = "Engine is mocked for this test"
+
+    with (
+        mock.patch(
+            "vllm.v1.engine.llm_engine.LLMEngine.add_request",
+            side_effect=MockEngineException(mock_err),
+        ) as mock_add_request,
+        vllm_runner(**vllm_runner_kwargs) as vllm_model,
+    ):
+        # Die once we actually submit the request to the engine
+        with pytest.raises(MockEngineException):
+            vllm_model.llm.generate(prompts=AUDIO_PROMPT)
+
+        # Then check to make sure the submitted lora request
+        # and text prompt were zipped together correctly
+        engine_args, engine_kwargs = mock_add_request.call_args
+        assert engine_kwargs["lora_request"] is None
+        assert engine_kwargs["prompt_text"] == AUDIO_PROMPT
diff --git a/vllm/entrypoints/llm.py b/vllm/entrypoints/llm.py
index b0786bd355aa6..7421eb8b8abc9 100644
--- a/vllm/entrypoints/llm.py
+++ b/vllm/entrypoints/llm.py
@@ -466,7 +466,7 @@ class LLM:
         ):
             return lora_request
 
-        if not isinstance(prompts, Sequence):
+        if not isinstance(prompts, Sequence) or isinstance(prompts, str):
             prompts = [prompts]
 
         optional_loras = (