diff --git a/docs/design/cuda_graphs_multimodal.md b/docs/design/cuda_graphs_multimodal.md
index 1fb5c2ba6514..5a9edc1ad93e 100644
--- a/docs/design/cuda_graphs_multimodal.md
+++ b/docs/design/cuda_graphs_multimodal.md
@@ -82,6 +82,7 @@ Models opt-in to encoder CUDA Graphs by implementing the [SupportsEncoderCudaGra
 
 | Architecture | Models | CG for Image | CG for Video |
 | ------------ | ------ | ------------ | ------------ |
+| `InternVLChatModel` | `InternVL3.5`, `InternVL3`, `InternVL2.5`, `InternVL2` | ✅︎ | ✅︎ |
 | `Qwen2VLForConditionalGeneration` | `Qwen2-VL` | ✅︎ | ✅︎ |
 | `Qwen2_5_VLForConditionalGeneration` | `Qwen2.5-VL` | ✅︎ | ✅︎ |
 | `Qwen3VLForConditionalGeneration` | `Qwen3-VL` | ✅︎ | ✅︎ |
diff --git a/examples/generate/multimodal/vision_language_offline.py b/examples/generate/multimodal/vision_language_offline.py
index b4e34bd64382..4d47d9f8b453 100644
--- a/examples/generate/multimodal/vision_language_offline.py
+++ b/examples/generate/multimodal/vision_language_offline.py
@@ -2554,6 +2554,7 @@ def run_tarsier2(questions: list[str], modality: str) -> ModelRequestData:
 
 
 MODELS_SUPPORT_VIT_CUDA_GRAPH = [
+    "internvl_chat",
     "qwen2_5_vl",
     "qwen3_vl",
     "qwen3_vl_moe",
diff --git a/tests/models/multimodal/generation/test_vit_cudagraph.py b/tests/models/multimodal/generation/test_vit_cudagraph.py
index 18630e3559a3..cbdc5e878aed 100644
--- a/tests/models/multimodal/generation/test_vit_cudagraph.py
+++ b/tests/models/multimodal/generation/test_vit_cudagraph.py
@@ -43,6 +43,10 @@ def qwen_vl_chat_template(content: str) -> str:
     return f"<|im_start|>user\n{content}<|im_end|>\n<|im_start|>assistant\n"
 
 
+def internvl_chat_template(content: str) -> str:
+    return f"<|im_start|>user\n{content}<|im_end|>\n<|im_start|>assistant\n"
+
+
 def step3_vl_chat_template(content: str) -> str:
     return (
         "<｜begin▁of▁sentence｜> You are a helpful assistant.<|BOT|>user\n "
@@ -51,6 +55,17 @@ def step3_vl_chat_template(content: str) -> str:
 
 
 MODEL_CONFIGS: dict[str, VitCudagraphTestConfig] = {
+    "internvl": VitCudagraphTestConfig(
+        model="OpenGVLab/InternVL3-1B",
+        num_video_frames=8,
+        image_prompt=internvl_chat_template("<image>\nWhat is in this image?"),
+        video_prompt=internvl_chat_template(
+            "<video>\nDescribe this video in one sentence."
+        ),
+        needs_video_metadata=False,
+        vllm_runner_kwargs={"trust_remote_code": True},
+        marks=[pytest.mark.core_model],
+    ),
     "qwen2_5_vl": VitCudagraphTestConfig(
         model="Qwen/Qwen2.5-VL-3B-Instruct",
         image_prompt=qwen_vl_chat_template(
diff --git a/vllm/model_executor/models/internvl.py b/vllm/model_executor/models/internvl.py
index f3918e302b47..d57614ea9802 100644
--- a/vllm/model_executor/models/internvl.py
+++ b/vllm/model_executor/models/internvl.py
@@ -10,7 +10,7 @@
 from abc import abstractmethod
 from collections.abc import Iterable, Mapping, Sequence
 from functools import cached_property
-from typing import Annotated, Literal, TypeAlias, TypeVar
+from typing import Annotated, Any, Literal, TypeAlias, TypeVar
 
 import torch
 import torch.nn as nn
@@ -55,6 +55,7 @@
 
 from .interfaces import (
     MultiModalEmbeddings,
+    SupportsEncoderCudaGraph,
     SupportsLoRA,
     SupportsMultiModal,
     SupportsPP,
@@ -543,7 +544,13 @@ def _get_prompt_updates(
     info=InternVLProcessingInfo,
     dummy_inputs=InternVLDummyInputsBuilder,
 )
-class InternVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA):
+class InternVLChatModel(
+    nn.Module,
+    SupportsMultiModal,
+    SupportsPP,
+    SupportsLoRA,
+    SupportsEncoderCudaGraph,
+):
     supports_encoder_tp_data = True
 
     @classmethod
@@ -924,3 +931,160 @@ def get_num_mm_connector_tokens(self, num_vision_tokens: int) -> int:
 
         num_patches = num_vision_tokens // (self.patch_tokens + 1)
         return num_patches * self.num_image_token
+
+    # -- SupportsEncoderCudaGraph protocol methods --
+
+    def get_encoder_cudagraph_config(self):
+        from vllm.v1.worker.encoder_cudagraph_defs import EncoderCudaGraphConfig
+
+        return EncoderCudaGraphConfig(
+            modalities=["image", "video"],
+            # InternVision uses standard ViT attention (no rotary embeddings,
+            # no variable-length sequence metadata), so the only graph-recorded
+            # buffer is pixel_values_flat itself.
+            buffer_keys=["pixel_values_flat"],
+            out_hidden_size=self.config.text_config.hidden_size,
+        )
+
+    def get_input_modality(
+        self,
+        mm_kwargs: dict[str, Any],
+    ) -> str:
+        if "pixel_values_flat" in mm_kwargs:
+            return "image"
+        return "video"
+
+    def get_encoder_cudagraph_budget_range(
+        self,
+        vllm_config: "VllmConfig",
+    ) -> tuple[int, int]:
+        # Min: 1 tile → num_image_token output tokens.
+        min_budget = self.num_image_token
+        max_budget = min(
+            vllm_config.scheduler_config.max_num_batched_tokens,
+            vllm_config.model_config.max_model_len,
+        )
+        return (min_budget, max_budget)
+
+    def _get_internvl_patches_list(
+        self,
+        mm_kwargs: dict[str, Any],
+    ) -> list[int]:
+        """Return per-item tile counts as a plain list of ints."""
+        if self.get_input_modality(mm_kwargs) == "image":
+            patches = mm_kwargs.get("image_num_patches", [])
+        else:
+            patches = mm_kwargs.get("video_num_patches", [])
+        if isinstance(patches, torch.Tensor):
+            return patches.tolist()
+        return [int(n) for n in patches]
+
+    def get_encoder_cudagraph_item_specs(
+        self,
+        mm_kwargs: dict[str, Any],
+    ):
+        from vllm.v1.worker.encoder_cudagraph_defs import EncoderItemSpec
+
+        return [
+            EncoderItemSpec(
+                input_size=n,
+                output_tokens=n * self.num_image_token,
+            )
+            for n in self._get_internvl_patches_list(mm_kwargs)
+        ]
+
+    def select_encoder_cudagraph_items(
+        self,
+        mm_kwargs: dict[str, Any],
+        indices: list[int],
+    ) -> dict[str, Any]:
+        modality = self.get_input_modality(mm_kwargs)
+        pv_key = (
+            "pixel_values_flat" if modality == "image" else "pixel_values_flat_video"
+        )
+        patches_key = (
+            "image_num_patches" if modality == "image" else "video_num_patches"
+        )
+
+        pixel_values = mm_kwargs[pv_key]
+        patches_list = self._get_internvl_patches_list(mm_kwargs)
+
+        if len(indices) == 0:
+            return {pv_key: pixel_values[:0], patches_key: []}
+
+        # Compute cumulative tile offsets for slicing pixel_values.
+        cum_patches = [0]
+        for n in patches_list:
+            cum_patches.append(cum_patches[-1] + n)
+
+        selected_pv = torch.cat(
+            [pixel_values[cum_patches[i] : cum_patches[i + 1]] for i in indices]
+        )
+        selected_patches = [patches_list[i] for i in indices]
+
+        return {pv_key: selected_pv, patches_key: selected_patches}
+
+    def prepare_encoder_cudagraph_capture_inputs(
+        self,
+        token_budget: int,
+        max_batch_size: int,
+        max_frames_per_batch: int,
+        device: torch.device,
+        dtype: torch.dtype,
+    ):
+        from vllm.v1.worker.encoder_cudagraph_defs import (
+            EncoderCudaGraphCaptureInputs,
+        )
+
+        # Size the buffer to hold the maximum possible tiles for this budget.
+        total_tiles = max(token_budget // self.num_image_token, 1)
+        image_size = self.config.vision_config.image_size
+
+        dummy_pixel_values = torch.randn(
+            total_tiles, 3, image_size, image_size, device=device, dtype=dtype
+        )
+
+        return EncoderCudaGraphCaptureInputs(
+            values={"pixel_values_flat": dummy_pixel_values},
+        )
+
+    def prepare_encoder_cudagraph_replay_buffers(
+        self,
+        mm_kwargs: dict[str, Any],
+        max_batch_size: int,
+        max_frames_per_batch: int,
+    ):
+        from vllm.v1.worker.encoder_cudagraph_defs import (
+            EncoderCudaGraphReplayBuffers,
+        )
+
+        modality = self.get_input_modality(mm_kwargs)
+        pv_key = (
+            "pixel_values_flat" if modality == "image" else "pixel_values_flat_video"
+        )
+        return EncoderCudaGraphReplayBuffers(
+            values={"pixel_values_flat": mm_kwargs[pv_key]},
+        )
+
+    def encoder_cudagraph_forward(
+        self,
+        values: dict[str, torch.Tensor],
+    ) -> torch.Tensor:
+        # The graph is always captured with pixel_values_flat as the input
+        # buffer. During video replay the manager copies video tiles into
+        # this same buffer before calling graph.replay(), so we always read
+        # from pixel_values_flat here.
+        pixel_values = values["pixel_values_flat"]
+        out = self.extract_feature(pixel_values)  # [N, num_image_token, H]
+        return out.view(-1, self.config.text_config.hidden_size)
+
+    def encoder_eager_forward(
+        self,
+        mm_kwargs: dict[str, Any],
+    ) -> torch.Tensor:
+        if self.get_input_modality(mm_kwargs) == "image":
+            pixel_values = mm_kwargs["pixel_values_flat"]
+        else:
+            pixel_values = mm_kwargs["pixel_values_flat_video"]
+        out = self.extract_feature(pixel_values)  # [N, num_image_token, H]
+        return out.view(-1, self.config.text_config.hidden_size)