fix(vlm): materialize Qwen3-VL features on the vision device (#31596)

python/sglang/srt/models/qwen3_vl.py

@@ -68,7 +68,10 @@ from sglang.srt.models.utils import (
     WeightsMapper,
     compute_cu_seqlens_from_grid_numpy,
 )
-from sglang.srt.multimodal.mm_utils import run_dp_sharded_mrope_vision_model
+from sglang.srt.multimodal.mm_utils import (
+    materialize_multimodal_features,
+    run_dp_sharded_mrope_vision_model,
+)
 from sglang.srt.multimodal.vit_cuda_graph_runner import ViTCudaGraphRunner
 from sglang.srt.runtime_context import get_parallel, get_server_args
 from sglang.srt.utils import (
@@ -1330,9 +1333,10 @@ class Qwen3VLForConditionalGeneration(nn.Module):
         return pattern.pad_input_tokens(input_ids, mm_inputs)
 
     def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
-        # in qwen-vl, last dim is the same
+        pixel_values = materialize_multimodal_features(
-        pixel_values = torch.cat([item.feature for item in items], dim=0).type(
+            [item.feature for item in items],
-            self.visual.dtype
+            device=self.visual.device,
+            dtype=self.visual.dtype,
         )
         image_grid_thw = torch.concat([item.image_grid_thw for item in items], dim=0)
         assert pixel_values.dim() == 2, pixel_values.dim()
@@ -1349,9 +1353,10 @@ class Qwen3VLForConditionalGeneration(nn.Module):
             return self.visual(pixel_values, grid_thw=image_grid_thw)
 
     def get_video_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
-        # in qwen-vl, last dim is the same
+        pixel_values = materialize_multimodal_features(
-        pixel_values = torch.cat([item.feature for item in items], dim=0).type(
+            [item.feature for item in items],
-            self.visual.dtype
+            device=self.visual.device,
+            dtype=self.visual.dtype,
         )
         video_grid_thw = torch.concat([item.video_grid_thw for item in items], dim=0)
         assert pixel_values.dim() == 2, pixel_values.dim()

test/registered/unit/models/test_qwen3_vl_feature_materialization.py

@@ -0,0 +1,75 @@
+"""Regression tests for Qwen3-VL multimodal feature materialization."""
+
+import unittest
+from types import SimpleNamespace
+
+import torch
+
+from sglang.srt.models.qwen3_vl import Qwen3VLForConditionalGeneration
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+
+class _RecordingVisual:
+    device = torch.device("meta")
+    dtype = torch.bfloat16
+
+    def __init__(self):
+        self.pixel_values = None
+        self.grid_thw = None
+
+    def __call__(self, pixel_values, *, grid_thw):
+        self.pixel_values = pixel_values
+        self.grid_thw = grid_thw
+        return pixel_values
+
+
+class TestQwen3VLFeatureMaterialization(CustomTestCase):
+    def test_image_features_are_packed_on_the_visual_device(self):
+        visual = _RecordingVisual()
+        model = SimpleNamespace(visual=visual, use_data_parallel=False)
+        items = [
+            SimpleNamespace(
+                feature=torch.ones(2, 3),
+                image_grid_thw=torch.tensor([[1, 1, 2]]),
+            ),
+            SimpleNamespace(
+                feature=torch.ones(1, 3),
+                image_grid_thw=torch.tensor([[1, 1, 1]]),
+            ),
+        ]
+        output = Qwen3VLForConditionalGeneration.get_image_feature(model, items)
+
+        self.assertIs(visual.pixel_values, output)
+        self.assertEqual(output.shape, (3, 3))
+        self.assertEqual(output.device, visual.device)
+        self.assertEqual(output.dtype, visual.dtype)
+
+    def test_video_features_are_packed_on_the_visual_device(self):
+        visual = _RecordingVisual()
+        model = SimpleNamespace(visual=visual, use_data_parallel=False)
+        items = [
+            SimpleNamespace(
+                feature=torch.ones(3, 4),
+                video_grid_thw=torch.tensor([[1, 1, 3]]),
+            ),
+            SimpleNamespace(
+                feature=torch.ones(2, 4),
+                video_grid_thw=torch.tensor([[1, 1, 2]]),
+            ),
+        ]
+        output = Qwen3VLForConditionalGeneration.get_video_feature(model, items)
+
+        self.assertIs(visual.pixel_values, output)
+        self.assertEqual(output.shape, (5, 4))
+        self.assertEqual(output.device, visual.device)
+        self.assertEqual(output.dtype, visual.dtype)
+        self.assertTrue(
+            torch.equal(visual.grid_thw, torch.tensor([[1, 1, 3], [1, 1, 2]]))
+        )
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)