[MM][CG] Support ViT CG for Qwen2.5-VL

tests/models/multimodal/generation/test_qwen2_5_vl.py

@@ -4,6 +4,7 @@
 import pytest
 
 from vllm.multimodal.video import sample_frames_from_video
+from vllm.platforms import current_platform
 
 from ....conftest import VIDEO_ASSETS
 
@@ -28,6 +29,15 @@ def qwen2_5_vl_chat_template(*query):
 )
 
 
+def _vision_encoder_cudagraph_compilation_config(*, cudagraph_mm_encoder: bool) -> dict:
+    # When `compilation_config` is passed, VllmRunner does not add defaults; keep
+    # decoder cudagraph sizes aligned with conftest and only toggle ViT paths.
+    return {
+        "cudagraph_capture_sizes": [4],
+        "cudagraph_mm_encoder": cudagraph_mm_encoder,
+    }
+
+
 @pytest.mark.core_model
 @pytest.mark.parametrize("model", models)
 @pytest.mark.parametrize("video_pruning_rate", [0.0, 0.75])
@@ -146,3 +156,63 @@ def test_qwen2_5_vl_evs_batched_videos(
 
             # Ensure the output is a string
             assert isinstance(output_text, str)
+
+
+@pytest.mark.core_model
+@pytest.mark.skipif(
+    not current_platform.is_cuda(), reason="ViT encoder CUDA graphs require CUDA"
+)
+def test_qwen2_5_vl_vision_transformer_cudagraph_matches_eager(
+    vllm_runner,
+    video_assets,
+) -> None:
+    """Greedy token ids should match with and without encoder CUDA graph replay.
+
+    Video EVS pruning must be off so the vision encoder is eligible for CUDA
+    graph capture (see ``get_encoder_cudagraph_config`` on Qwen2.5-VL).
+    """
+    model = models[0]
+    dtype = target_dtype
+    num_frames = 16
+    max_tokens = 64
+    # No EVS: ``video_pruning_rate`` must not be > 0, or video CUDA graphs
+    # are disabled and this test would not exercise the vision encoder path.
+    sampled_vids = [
+        sample_frames_from_video(asset.np_ndarrays, num_frames)
+        for asset in video_assets
+    ]
+    prompts = [VIDEO_PROMPTS[0]]
+    videos = [sampled_vids[0]]
+    # Encoder CUDA graphs add substantial graph-pool memory; cap concurrency and
+    # GPU fraction so init (decoder graphs + optional encoder graphs + KV) fits.
+    common_kwargs = dict(
+        runner="generate",
+        max_model_len=4000,
+        max_num_seqs=16,
+        gpu_memory_utilization=0.85,
+        dtype=dtype,
+        limit_mm_per_prompt={"video": 1},
+        video_pruning_rate=0.0,
+    )
+
+    with vllm_runner(
+        model,
+        **common_kwargs,
+        compilation_config=_vision_encoder_cudagraph_compilation_config(
+            cudagraph_mm_encoder=False
+        ),
+    ) as mm_eager:
+        eager_ids = mm_eager.generate_greedy(prompts, max_tokens, videos=videos)[0][0]
+
+    with vllm_runner(
+        model,
+        **common_kwargs,
+        compilation_config=_vision_encoder_cudagraph_compilation_config(
+            cudagraph_mm_encoder=True
+        ),
+    ) as mm_cudagraph:
+        graph_ids = mm_cudagraph.generate_greedy(prompts, max_tokens, videos=videos)[0][
+            0
+        ]
+
+    assert eager_ids == graph_ids