[Fix] Carry the backend on Kimi-K3 deferred preprocessing configs (#34766)

test/registered/unit/disaggregation/test_kimi_k3_encoder_mode.py

@@ -188,8 +188,8 @@ def test_kimi_k3_epd_preprocess_preserves_raw_per_image_items():
         assert item.hash is not None
         assert item.pad_value is not None
         deferred = item.model_specific_data[DEFERRED_PREPROCESSING_KEY]
-        assert deferred["image_mean"] == [0.5, 0.5, 0.5]
-        assert deferred["image_std"] == [0.5, 0.5, 0.5]
+        assert deferred.image_mean == [0.5, 0.5, 0.5]
+        assert deferred.image_std == [0.5, 0.5, 0.5]
 
 
 def test_kimi_k3_epd_model_preprocessor_receives_image_processor():
@@ -289,7 +289,7 @@ def test_kimi_k3_epd_default_cpu_materialization_is_owner_only_and_exact():
     assert len(processor.calls[0]) == 1
     assert processor.calls[0][0]["image"].getpixel((0, 0)) == (11, 0, 0)
     assert torch.all(materialized == 11)
-    assert items[0].model_specific_data[DEFERRED_PREPROCESSING_KEY]["backend"] == "cpu"
+    assert items[0].model_specific_data[DEFERRED_PREPROCESSING_KEY].backend == "cpu"
 
 
 def test_encoder_preprocess_materializes_only_local_size_balanced_items():

test/registered/unit/models/test_kimi_k25.py

@@ -1,6 +1,7 @@
 """CPU coverage for Kimi-K2.5/K2.7 encoder-DP wiring."""
 
 import asyncio
+import functools
 from types import SimpleNamespace
 from unittest.mock import AsyncMock, Mock, patch
 
@@ -690,6 +691,7 @@ def test_kimi_k3_epd_rebuild_uses_the_same_media_contract():
 def test_kimi_k3_cpu_transport_defers_gpu_preprocessing():
     from sglang.srt.multimodal.kimi_k3_image_processing import (
         DEFERRED_PREPROCESSING_KEY,
+        KimiK3DeferredPreprocessing,
     )
 
     processor = object.__new__(KimiK3ImageProcessor)
@@ -717,11 +719,13 @@ def test_kimi_k3_cpu_transport_defers_gpu_preprocessing():
                         "pad_height": 2,
                     },
                 ],
-                {
-                    "image_mean": [0.5, 0.5, 0.5],
-                    "image_std": [0.5, 0.5, 0.5],
-                    "transparent_bg_config": None,
-                },
+                functools.partial(
+                    KimiK3DeferredPreprocessing,
+                    backend="gpu",
+                    image_mean=[0.5, 0.5, 0.5],
+                    image_std=[0.5, 0.5, 0.5],
+                    transparent_bg_config=None,
+                ),
             )
         ),
     )
@@ -751,10 +755,13 @@ def test_kimi_k3_cpu_transport_defers_gpu_preprocessing():
     ]
     assert all(item.hash is not None for item in output.mm_items)
     assert all(item.pad_value is not None for item in output.mm_items)
-    assert all(
-        DEFERRED_PREPROCESSING_KEY in item.model_specific_data
-        for item in output.mm_items
-    )
+    deferred = [
+        item.model_specific_data[DEFERRED_PREPROCESSING_KEY] for item in output.mm_items
+    ]
+    # The staged features are CHW uint8, so the config has to route them to the
+    # GPU arm of `materialize_item_features`.
+    assert [config.backend for config in deferred] == ["gpu", "gpu"]
+    assert [config.resize_config["new_width"] for config in deferred] == [4, 2]
 
 
 @pytest.mark.parametrize(

test/registered/unit/models/test_kimi_k3_vision.py

@@ -509,6 +509,7 @@ def test_kimi_k3_preprocesses_only_dp_owner_images(monkeypatch):
     from sglang.srt.models.kimi_k3 import KimiK3ForConditionalGeneration
     from sglang.srt.multimodal.kimi_k3_image_processing import (
         DEFERRED_PREPROCESSING_KEY,
+        KimiK3DeferredPreprocessing,
     )
 
     model = KimiK3ForConditionalGeneration.__new__(KimiK3ForConditionalGeneration)
@@ -517,20 +518,19 @@ def test_kimi_k3_preprocesses_only_dp_owner_images(monkeypatch):
     model.vision_tower = _K3TowerStub()
     model.mm_projector = lambda image_embeds: image_embeds
 
-    deferred_config = {
-        "backend": "gpu",
-        "feature_layout": "chw",
-        "image_mean": [0.5, 0.5, 0.5],
-        "image_std": [0.5, 0.5, 0.5],
-        "transparent_bg_config": None,
-        "resize_config": {
+    deferred_config = KimiK3DeferredPreprocessing(
+        backend="gpu",
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+        transparent_bg_config=None,
+        resize_config={
             "num_tokens": 1,
             "new_width": 2,
             "new_height": 2,
             "pad_width": 0,
             "pad_height": 0,
         },
-    }
+    )
     items = [
         MultimodalDataItem(
             modality=Modality.IMAGE,