[BugFix][Multimodal] Pick worst-case DeepSeek-V4 VL dummy image size (#59271) (#59373)

Signed-off-by: Yida Weng <84164537+YidaWeng@users.noreply.github.com>
Co-authored-by: Grok <grok@x.ai>

tests/models/multimodal/processing/test_deepseek_v4_vl.py

@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""DeepSeek-V4 VL processing helpers."""
+
+from PIL import Image
+
+from vllm.models.deepseek_v4.common.mm_preprocess import (
+    DeepseekV4VLProcessingInfo,
+    DeepseekV4VLProcessor,
+    build_image_block_pad_free,
+)
+from vllm.transformers_utils.configs.deepseek_v4 import DeepseekV4Config
+
+
+class _ConfigCtx:
+    def __init__(self, config: DeepseekV4Config) -> None:
+        self._config = config
+
+    def get_hf_config(self, *args, **kwargs):
+        return self._config
+
+
+def _feature_counts(
+    processor: DeepseekV4VLProcessor, width: int, height: int
+) -> tuple[int, int]:
+    out = processor(images=[Image.new("RGB", (width, height))])
+    n_llm_h, n_llm_w = out["llm_grid"][0].tolist()
+    types, _ = build_image_block_pad_free(n_llm_h, n_llm_w)
+    return len(types), out["patches"].shape[0]
+
+
+def test_get_image_size_with_most_features_beats_wide_real_image():
+    """Dummy size must maximize tokens/patches within the budget (#59271).
+
+    A square under-counts because each aligner row pays an IMAGE_NEW_LINE.
+    """
+    config = DeepseekV4Config(vision_n_layers=24)
+    info = DeepseekV4VLProcessingInfo(_ConfigCtx(config))
+    processor = DeepseekV4VLProcessor(config)
+
+    dummy = info.get_image_size_with_most_features()
+    dummy_features = _feature_counts(processor, dummy.width, dummy.height)
+    wide_features = _feature_counts(processor, 1932, 336)
+
+    assert dummy == (1932, 336)
+    assert wide_features <= dummy_features
+
+
+def test_get_image_size_with_most_features_respects_max_wh_ratio():
+    config = DeepseekV4Config(vision_n_layers=24)
+    info = DeepseekV4VLProcessingInfo(_ConfigCtx(config))
+    dummy = info.get_image_size_with_most_features()
+    assert dummy.width / dummy.height <= config.vision_max_wh_ratio

vllm/models/deepseek_v4/common/mm_preprocess.py

@@ -354,15 +354,34 @@ class DeepseekV4VLProcessingInfo(BaseProcessingInfo):
         hf_config = self.get_hf_config()
         patch_size = hf_config.vision_patch_size
         downsample_ratio = hf_config.vision_downsample_ratio
-        # A square maximizes the ViT patch count (area) within the token
-        # budget; solve the budget-derived size directly to keep the dummy
-        # image small.
+        max_wh_ratio = hf_config.vision_max_wh_ratio
+        # Every aligner row costs an extra IMAGE_NEW_LINE token, so height is
+        # more expensive than width and a square is *not* optimal: search the
+        # LLM grids that fit the budget for the one with the most ViT patches
+        # (ties broken by the longer sentinel block).
         budget = hf_config.vision_max_n_token - (COMPRESS_PAD_TO - 1)
-        side = budget * patch_size * downsample_ratio
-        _, _, best_h, best_w, _ = solve_resize_ratio(
-            side, side, patch_size, downsample_ratio, budget
-        )
-        return ImageSize(width=best_w, height=best_h)
+        step = patch_size * downsample_ratio
+        best_grid = (1, 1)
+        best_key = (0, 0)
+        for n_llm_h in range(1, budget):
+            if (
+                grid_tokens(n_llm_h * step, step, patch_size, downsample_ratio)[2]
+                > budget
+            ):
+                break
+            for n_llm_w in range(1, budget):
+                if max_wh_ratio is not None and n_llm_w > n_llm_h * max_wh_ratio:
+                    break
+                num_tokens = grid_tokens(
+                    n_llm_h * step, n_llm_w * step, patch_size, downsample_ratio
+                )[2]
+                if num_tokens > budget:
+                    break
+                key = (n_llm_h * n_llm_w, num_tokens)
+                if key > best_key:
+                    best_key, best_grid = key, (n_llm_h, n_llm_w)
+        n_llm_h, n_llm_w = best_grid
+        return ImageSize(width=n_llm_w * step, height=n_llm_h * step)
 
 
 class DeepseekV4VLDummyInputsBuilder(