From a21cd9ed239b853bd587ffe3c9140fe68cd41f59 Mon Sep 17 00:00:00 2001
From: Isotr0py <mozf@mail2.sysu.edu.cn>
Date: Wed, 3 Dec 2025 18:05:10 +0800
Subject: [PATCH] [Bugfix] Fix incorrect `image_grid_thw` rank for HunyuanOCR
 from missing `merge_by_field_config=True` (#29950)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
---
 .../vision_language_multi_image.py            | 23 +++++++++++++++++++
 vllm/model_executor/models/hunyuan_vision.py  |  1 +
 2 files changed, 24 insertions(+)

diff --git a/examples/offline_inference/vision_language_multi_image.py b/examples/offline_inference/vision_language_multi_image.py
index 2193b1ca9cf4..560ca768d1a6 100755
--- a/examples/offline_inference/vision_language_multi_image.py
+++ b/examples/offline_inference/vision_language_multi_image.py
@@ -309,6 +309,28 @@ def load_h2ovl(question: str, image_urls: list[str]) -> ModelRequestData:
     )
 
 
+# HunyuanOCR
+def load_hunyuan_vl(question: str, image_urls: list[str]) -> ModelRequestData:
+    model_name = "tencent/HunyuanOCR"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=8192,
+        limit_mm_per_prompt={"image": len(image_urls)},
+    )
+
+    placeholder = (
+        "<｜hy_place▁holder▁no▁100｜><｜hy_place▁holder▁no▁102｜><｜hy_place▁holder▁no▁101｜>"  # noqa: E501
+    ) * len(image_urls)
+    prompt = f"<｜hy_begin▁of▁sentence｜>{placeholder}{question}<｜hy_User｜>"
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompt=prompt,
+        image_data=[fetch_image(url) for url in image_urls],
+    )
+
+
 def load_hyperclovax_seed_vision(
     question: str, image_urls: list[str]
 ) -> ModelRequestData:
@@ -1322,6 +1344,7 @@ model_example_map = {
     "deepseek_ocr": load_deepseek_ocr,
     "gemma3": load_gemma3,
     "h2ovl_chat": load_h2ovl,
+    "hunyuan_vl": load_hunyuan_vl,
     "hyperclovax_seed_vision": load_hyperclovax_seed_vision,
     "idefics3": load_idefics3,
     "interns1": load_interns1,
diff --git a/vllm/model_executor/models/hunyuan_vision.py b/vllm/model_executor/models/hunyuan_vision.py
index 2950db571e6e..6537b6df876a 100644
--- a/vllm/model_executor/models/hunyuan_vision.py
+++ b/vllm/model_executor/models/hunyuan_vision.py
@@ -785,6 +785,7 @@ class HunYuanVLForConditionalGeneration(
     SupportsQuant,
     SupportsXDRoPE,
 ):
+    merge_by_field_config = True
     multimodal_cpu_fields = {"image_grid_thw"}
 
     # To ensure correct weight loading and mapping.