diff --git a/.github/requirements-test.txt b/.github/requirements-test.txt
index 50b8db80..482ec386 100644
--- a/.github/requirements-test.txt
+++ b/.github/requirements-test.txt
@@ -2,6 +2,7 @@ codetiming
 datasets
 pillow
 pytest
+qwen-vl-utils
 ray[default]
 ruff
 tensordict
diff --git a/verl/trainer/config.py b/verl/trainer/config.py
index 0ded93ea..66566811 100644
--- a/verl/trainer/config.py
+++ b/verl/trainer/config.py
@@ -38,6 +38,7 @@ class DataConfig:
     prompt_key: str = "prompt"
     answer_key: str = "answer"
     image_key: str = "images"
+    video_key: str = "videos"
     image_dir: Optional[str] = None
     max_prompt_length: int = 512
     max_response_length: int = 512
diff --git a/verl/trainer/data_loader.py b/verl/trainer/data_loader.py
index 6491856d..e3a66768 100644
--- a/verl/trainer/data_loader.py
+++ b/verl/trainer/data_loader.py
@@ -32,6 +32,7 @@ def create_dataloader(config: DataConfig, tokenizer: PreTrainedTokenizer, proces
         answer_key=config.answer_key,
         image_key=config.image_key,
         image_dir=config.image_dir,
+        video_key=config.video_key,
         max_prompt_length=config.max_prompt_length,
         truncation="right",
         format_prompt=config.format_prompt,
diff --git a/verl/utils/dataset.py b/verl/utils/dataset.py
index e8f909b3..e1b174ad 100644
--- a/verl/utils/dataset.py
+++ b/verl/utils/dataset.py
@@ -30,6 +30,7 @@
 from ..models.transformers.qwen2_vl import get_rope_index
 from . import torch_functional as VF
 
+from qwen_vl_utils.vision_process import fetch_video
 
 def collate_fn(features: List[Dict[str, Any]]) -> Dict[str, Any]:
     tensors = defaultdict(list)
@@ -91,6 +92,7 @@ def __init__(
         answer_key: str = "answer",
         image_key: str = "images",
         image_dir: Optional[str] = None,
+        video_key: str = "videos",
         max_prompt_length: int = 1024,
         truncation: str = "error",
         format_prompt: Optional[str] = None,
@@ -104,6 +106,7 @@ def __init__(
         self.answer_key = answer_key
         self.image_key = image_key
         self.image_dir = image_dir
+        self.video_key = video_key
         self.max_prompt_length = max_prompt_length
         self.truncation = truncation
         self.min_pixels = min_pixels
@@ -152,6 +155,16 @@ def _build_messages(self, example: Dict[str, Any]) -> List[Dict[str, Any]]:
                 if content:
                     content_list.append({"type": "text", "text": content})
 
+            return [{"role": "user", "content": content_list}]
+        elif self.video_key in example:
+            content_list = []
+            for i, content in enumerate(prompt_str.split("<video>")):
+                if i != 0:
+                    content_list.append({"type": "video"})
+
+                if content:
+                    content_list.append({"type": "text", "text": content})
+
             return [{"role": "user", "content": content_list}]
         else:
             return [{"role": "user", "content": prompt_str}]
@@ -193,6 +206,18 @@ def __getitem__(self, index):
             input_ids = model_inputs.pop("input_ids")[0]
             attention_mask = model_inputs.pop("attention_mask")[0]
             example["multi_modal_data"] = {"images": images}
+        elif self.video_key in example:
+            prompt = self.processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
+            video = example.pop(self.video_key)
+            vision_info = {"video": video, "min_pixels": self.min_pixels, "max_pixels": self.max_pixels, "fps": 25}
+            video_input, video_sample_fps = fetch_video(vision_info, return_video_sample_fps=True)
+
+            model_inputs = self.processor(
+                videos=video_input, text=[prompt], add_special_tokens=False, return_tensors="pt"
+            )
+            input_ids = model_inputs.pop("input_ids")[0]
+            attention_mask = model_inputs.pop("attention_mask")[0]
+            example["multi_modal_data"] = {"video": video_input}
         else:
             prompt = self.tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
             model_inputs = self.tokenizer([prompt], add_special_tokens=False, return_tensors="pt")
@@ -204,7 +229,8 @@ def __getitem__(self, index):
             position_ids = get_rope_index(
                 self.processor,
                 input_ids=input_ids,
-                image_grid_thw=model_inputs.get("image_grid_thw"),
+                image_grid_thw=model_inputs.get("image_grid_thw", None),
+                video_grid_thw=model_inputs.get("video_grid_thw", None),
                 attention_mask=attention_mask,
             )  # (3, seq_length)
         else:
diff --git a/verl/workers/fsdp_workers.py b/verl/workers/fsdp_workers.py
index 9de09519..f24663f0 100644
--- a/verl/workers/fsdp_workers.py
+++ b/verl/workers/fsdp_workers.py
@@ -455,17 +455,21 @@ def _process_multi_modal_inputs(self, data: DataProto):
             max_pixels = data.meta_info["max_pixels"]
             batch_multi_modal_inputs = []
             for multi_modal_data in data.non_tensor_batch["multi_modal_data"]:
-                images = []
-                for image in multi_modal_data["images"]:
-                    images.append(process_image(image, min_pixels=min_pixels, max_pixels=max_pixels))
-
-                if len(images) != 0:
+                if "images" in multi_modal_data:
                     # it's necessary to add `dict` to properly convert batch features to dict
                     # otherwise the batch features will be converted to dict keys
                     # see https://github.com/hiyouga/EasyR1/pull/339
+                    images = []
+                    for image in multi_modal_data["images"]:
+                        images.append(process_image(image, min_pixels=min_pixels, max_pixels=max_pixels))
                     multi_modal_inputs = dict(self.processor.image_processor(images=images, return_tensors="pt"))
                     multi_modal_inputs = {k: v.to(torch.cuda.current_device()) for k, v in multi_modal_inputs.items()}
                     batch_multi_modal_inputs.append(multi_modal_inputs)
+                elif "video" in multi_modal_data:
+                    video = multi_modal_data["video"]
+                    multi_modal_inputs = dict(self.processor.image_processor(images=None, videos=video, return_tensors="pt"))
+                    multi_modal_inputs = {k: v.to(torch.cuda.current_device()) for k, v in multi_modal_inputs.items()}
+                    batch_multi_modal_inputs.append(multi_modal_inputs)
                 else:
                     batch_multi_modal_inputs.append({})
 
diff --git a/verl/workers/rollout/vllm_rollout_spmd.py b/verl/workers/rollout/vllm_rollout_spmd.py
index 9e2610cf..4f98e318 100644
--- a/verl/workers/rollout/vllm_rollout_spmd.py
+++ b/verl/workers/rollout/vllm_rollout_spmd.py
@@ -52,12 +52,16 @@ def _get_logit_bias(processor: Optional[ProcessorMixin]) -> Optional[Dict[int, f
 def _process_multi_modal_data(multi_modal_data: Dict[str, Any], min_pixels: int, max_pixels: int) -> Dict[str, Any]:
     # may convert image path to image object
     # TODO: add video
-    images = []
-    for image in multi_modal_data["images"]:
-        images.append(process_image(image, min_pixels=min_pixels, max_pixels=max_pixels))
-
-    if len(images) != 0:
-        return {"image": images}
+    
+    if "images" in multi_modal_data:
+        images = []
+        for image in multi_modal_data["images"]:
+            images.append(process_image(image, min_pixels=min_pixels, max_pixels=max_pixels))
+            return {"image": images}
+    elif "video" in multi_modal_data:
+        # for image in multi_modal_data["video"]:
+        #     images.append(process_image(image, min_pixels=min_pixels, max_pixels=max_pixels))
+        return {"video": multi_modal_data["video"]}
 
     return None