From 0da126676e38f5dc1b02b2c4f48dc85f90f5a79f Mon Sep 17 00:00:00 2001
From: "Kevin H. Luu" <khluu000@gmail.com>
Date: Sun, 27 Sep 2026 21:52:51 -0700
Subject: [PATCH] [Bugfix] Fix the two multimodal root tests that fail on main
 (OpenPangu-VL embed merge, MiMo sink test fixture) (#58900)

Signed-off-by: khluu <khluu000@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
---
 tests/models/multimodal/test_mimo_v2_omni.py | 21 +++-----------------
 vllm/model_executor/models/openpangu_vl.py   | 21 +++++++++++---------
 2 files changed, 15 insertions(+), 27 deletions(-)

diff --git a/tests/models/multimodal/test_mimo_v2_omni.py b/tests/models/multimodal/test_mimo_v2_omni.py
index d4e1d9ec18..deccace50d 100644
--- a/tests/models/multimodal/test_mimo_v2_omni.py
+++ b/tests/models/multimodal/test_mimo_v2_omni.py
@@ -5,13 +5,7 @@
 import pytest
 import torch
 
-from tests.utils import ensure_current_vllm_config
-from vllm.distributed.parallel_state import (
-    init_distributed_environment,
-    initialize_model_parallel,
-)
 from vllm.platforms import current_platform
-from vllm.utils.network_utils import get_open_port
 
 EMBED_DIM = 256
 NUM_HEADS = 4
@@ -21,20 +15,11 @@ WINDOW = 8
 SEQ_LENS = [5, 37]
 
 
-@pytest.fixture(scope="module")
-def vision_attn_env():
-    init_distributed_environment(
-        world_size=1,
-        rank=0,
-        local_rank=0,
-        distributed_init_method=f"tcp://127.0.0.1:{get_open_port()}",
-        backend="nccl",
-    )
+@pytest.fixture
+def vision_attn_env(dist_init):
     default_dtype = torch.get_default_dtype()
     torch.set_default_dtype(torch.bfloat16)
-    with ensure_current_vllm_config():
-        initialize_model_parallel(tensor_model_parallel_size=1)
-        yield
+    yield
     torch.set_default_dtype(default_dtype)
 
 
diff --git a/vllm/model_executor/models/openpangu_vl.py b/vllm/model_executor/models/openpangu_vl.py
index b7d5c81032..8a62aa6ba6 100644
--- a/vllm/model_executor/models/openpangu_vl.py
+++ b/vllm/model_executor/models/openpangu_vl.py
@@ -1055,15 +1055,18 @@ class OpenPanguVLForConditionalGeneration(
         input_ids: torch.Tensor,
         multimodal_embeddings: MultiModalEmbeddings | None = None,
     ) -> torch.Tensor:
-        inputs_embeds = self.language_model.embed_input_ids(input_ids)
-        if multimodal_embeddings is not None:
-            inputs_embeds = self.embed_input_ids(  # type: ignore[call-overload]
-                input_ids,
-                inputs_embeds,
-                multimodal_embeddings,
-                [self.config.image_token_id, self.config.video_token_id],
-            )
-        return inputs_embeds
+        if multimodal_embeddings is None:
+            return self.language_model.embed_input_ids(input_ids)
+
+        mm_token_ids = input_ids.new_tensor(
+            [self.config.image_token_id, self.config.video_token_id]
+        )
+        is_multimodal = torch.isin(input_ids, mm_token_ids)
+        return self.embed_input_ids(
+            input_ids,
+            multimodal_embeddings,
+            is_multimodal=is_multimodal,
+        )
 
     def _process_image_input(
         self, image_input: OpenPanguVLImageInputs