[CI] Fix "test_vit_cudagraph_[image|video][step3_vl]" failure (#43082)

Signed-off-by: haosdent <haosdent@gmail.com>

tests/models/multimodal/generation/test_vit_cudagraph.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from dataclasses import dataclass, field
+from functools import partial
 
 import pytest
 
@@ -10,6 +11,7 @@ from vllm.platforms import current_platform
 
 from ....conftest import IMAGE_ASSETS, VIDEO_ASSETS
 from ....utils import create_new_process_for_each_test
+from ...utils import dummy_hf_overrides
 from .vlm_utils.builders import sample_frames_with_video_metadata
 
 
@@ -26,6 +28,7 @@ class VitCudagraphTestConfig:
     num_video_frames: int = 16
     needs_video_metadata: bool = False
     vllm_runner_kwargs: dict = field(default_factory=dict)
+    compilation_config_overrides: dict = field(default_factory=dict)
     marks: list = field(default_factory=list)
 
 
@@ -99,17 +102,36 @@ MODEL_CONFIGS: dict[str, VitCudagraphTestConfig] = {
     ),
     "step3_vl": VitCudagraphTestConfig(
         model="stepfun-ai/Step3-VL-10B",
+        modalities=["image"],
         image_prompt=step3_vl_chat_template("What is in this image?"),
-        video_prompt=None,
+        # Single bucket sized to cover the largest test image's output
+        # tokens (1152 > 1141 for cherry_blossom). The default auto-
+        # inferred range fans out into multiple power-of-2 buckets, each
+        # holding a full ViT capture pool.
+        compilation_config_overrides={
+            "encoder_cudagraph_token_budgets": [1152],
+        },
+        # Shrink to 1 text + 1 vision layer with random weights so the
+        # test runs on any CI GPU (incl. L4) and skips the 20 GiB weight
+        # download. The test only validates that encoder CG capture/
+        # replay functions correctly, not output quality.
+        vllm_runner_kwargs={
+            "load_format": "dummy",
+            "hf_overrides": partial(
+                dummy_hf_overrides,
+                model_arch="StepVLForConditionalGeneration",
+            ),
+        },
     ),
 }
 
 
-def get_compilation_config():
+def get_compilation_config(config: VitCudagraphTestConfig):
     return {
         "cudagraph_mm_encoder": True,
         "encoder_cudagraph_max_vision_items_per_batch": 1,
         "encoder_cudagraph_max_frames_per_batch": 16,
+        **config.compilation_config_overrides,
     }
 
 
@@ -141,7 +163,7 @@ def test_vit_cudagraph_image(model_id, vllm_runner, image_assets):
         max_model_len=config.max_model_len,
         max_num_seqs=config.max_num_seqs,
         limit_mm_per_prompt={"image": 1},
-        compilation_config=get_compilation_config(),
+        compilation_config=get_compilation_config(config),
         **config.vllm_runner_kwargs,
     ) as vllm_model:
         outputs = vllm_model.generate_greedy(
@@ -194,7 +216,7 @@ def test_vit_cudagraph_video(model_id, vllm_runner, video_assets):
         max_model_len=config.max_model_len,
         max_num_seqs=config.max_num_seqs,
         limit_mm_per_prompt={"video": 1},
-        compilation_config=get_compilation_config(),
+        compilation_config=get_compilation_config(config),
         **config.vllm_runner_kwargs,
     ) as vllm_model:
         outputs = vllm_model.generate_greedy(