[VLM] avoid CUDA placement on non-CUDA platforms (#38750)

python/sglang/srt/multimodal/processors/base_processor.py

@@ -23,6 +23,7 @@ import torch
 from PIL import Image
 from transformers import BaseImageProcessor
 
+from sglang.srt import platforms
 from sglang.srt.managers.schedule_batch import (
     Modality,
     MultimodalDataItem,
@@ -721,19 +722,23 @@ class BaseMultimodalProcessor(ABC):
         return processor, _tokenizer_of(processor)
 
     def _preprocessing_competes_with_the_scheduler(self) -> bool:
-        """Whether image preprocessing submits its work to the serving GPU.
+        """Whether image preprocessing contends with the serving accelerator.
 
-        The fast image processor runs inside the tokenizer process but on
+        The fast image processor runs inside the tokenizer process but may run on
-        ``cuda:{base_gpu_id}`` -- the device the scheduler serves from. A second
+        the same accelerator as the scheduler. A second preprocessing worker there
-        preprocessing worker there is one more competitor for that device rather
+        adds device contention rather than CPU preprocessing parallelism.
-        than added parallelism.
         """
         if _is_cpu or get_exec().deterministic.rl_on_policy_target is not None:
             return False
         if self.disable_fast_image_processor:
             return False
         image_processor = getattr(self._processor, "image_processor", None)
-        return isinstance(image_processor, BaseImageProcessor)
+        if not isinstance(image_processor, BaseImageProcessor):
+            return False
+        if _is_xpu or _is_npu:
+            return True
+        platform = platforms.current_platform
+        return platform.is_cuda_alike()
 
     def _resolve_auto_mm_processor_worker_num(self) -> int:
         """The worker count to use when the user did not ask for one.
@@ -769,9 +774,12 @@ class BaseMultimodalProcessor(ABC):
         if _is_xpu:
             return "xpu"
         if not _is_npu:
+            platform = platforms.current_platform
+            if not platform.is_cuda_alike():
+                return None
             # Per-worker placement travels as a constructor argument, and
             # this record is that argument.
-            return f"cuda:{server_args.base_gpu_id}"
+            return f"{platform.device_type}:{server_args.base_gpu_id}"
         if processor.__class__.__name__ == "MiniMaxVLProcessor":
             # MiniMax's image/video processors create 10-dim tensors during
             # patch extraction, exceeding the Ascend 8-dim limit; patch them

test/registered/unit/managers/test_mm_process_config.py

@@ -75,6 +75,21 @@ class TestMmProcessConfigValidation(CustomTestCase):
 class TestBaseProcessorConfigExtraction(CustomTestCase):
     """Verify BaseMultimodalProcessor.__init__ extracts configs from server_args."""
 
+    def _patch_platform(self, cuda_alike, device_type):
+        platforms = SimpleNamespace(
+            current_platform=SimpleNamespace(
+                is_cuda_alike=lambda: cuda_alike,
+                device_type=device_type,
+            )
+        )
+        return patch.multiple(
+            "sglang.srt.multimodal.processors.base_processor",
+            _is_cpu=False,
+            _is_xpu=False,
+            _is_npu=False,
+            platforms=platforms,
+        )
+
     def _make_processor(
         self,
         mm_process_config,
@@ -210,12 +225,23 @@ class TestBaseProcessorConfigExtraction(CustomTestCase):
         9.30 -> 4.02 req/s on GB300 for full-page images."""
         from transformers import BaseImageProcessor
 
-        proc = self._make_processor(
+        with self._patch_platform(cuda_alike=True, device_type="cuda"):
-            {}, image_processor=MagicMock(spec=BaseImageProcessor)
+            proc = self._make_processor(
-        )
+                {}, image_processor=MagicMock(spec=BaseImageProcessor)
+            )
         self.assertEqual(proc.mm_processor_worker_num, 1)
         self.assertIsNone(proc.mm_processor_executor)
 
+    def test_non_accelerator_fast_processor_gets_two_workers(self):
+        from transformers import BaseImageProcessor
+
+        with self._patch_platform(cuda_alike=False, device_type="custom"):
+            proc = self._make_processor(
+                {}, image_processor=MagicMock(spec=BaseImageProcessor)
+            )
+        self.assertEqual(proc.mm_processor_worker_num, 2)
+        self.assertIsNotNone(proc.mm_processor_executor)
+
     def test_explicit_request_overrides_the_path_decision(self):
         """The server argument wins: an operator who measured their own workload
         can still ask for concurrency on the GPU path."""
@@ -240,7 +266,10 @@ class TestBaseProcessorConfigExtraction(CustomTestCase):
             BaseMultimodalProcessor,
         )
 
-        with patch.object(BaseMultimodalProcessor, "auto_mm_processor_worker_num", 3):
+        with (
+            patch.object(BaseMultimodalProcessor, "auto_mm_processor_worker_num", 3),
+            self._patch_platform(cuda_alike=True, device_type="cuda"),
+        ):
             proc = self._make_processor(
                 {}, image_processor=MagicMock(spec=BaseImageProcessor)
             )

test/registered/unit/multimodal/test_processor_device_selection.py

@@ -51,10 +51,16 @@ class TestFastImageProcessorDevice(CustomTestCase):
         reset_context()
         self.addCleanup(reset_context)
 
-    def _device(self, processor, **platform):
+    def _device(self, processor, cuda_alike=True, device_type="cuda", **platform):
         flags = {"_is_cpu": False, "_is_xpu": False, "_is_npu": False}
         flags.update(platform)
-        with patch.multiple(BASE, **flags):
+        platforms = SimpleNamespace(
+            current_platform=SimpleNamespace(
+                is_cuda_alike=lambda: cuda_alike,
+                device_type=device_type,
+            )
+        )
+        with patch.multiple(BASE, platforms=platforms, **flags):
             return processor._fast_image_processor_device(_Processor())
 
     def test_device_follows_the_instance_base_gpu_id(self):
@@ -92,6 +98,23 @@ class TestFastImageProcessorDevice(CustomTestCase):
             device = processor._fast_image_processor_device(Glm4vProcessor())
         self.assertIsNone(device)
 
+    def test_platform_respects_cuda_compatibility(self):
+        processor = _make(base_gpu_id=3)
+        for cuda_alike, device_type, expected in (
+            (False, "other", None),
+            (True, "cuda", "cuda:3"),
+            (True, "musa", "musa:3"),
+        ):
+            with self.subTest(cuda_alike=cuda_alike, device_type=device_type):
+                self.assertEqual(
+                    self._device(
+                        processor,
+                        cuda_alike=cuda_alike,
+                        device_type=device_type,
+                    ),
+                    expected,
+                )
+
 
 class TestFastImageProcessorMemoryPool(CustomTestCase):
     def setUp(self):
@@ -161,6 +184,18 @@ class TestFastImageProcessorMemoryPool(CustomTestCase):
                 events.append("exit")
 
         with (
+            patch.multiple(
+                BASE,
+                _is_cpu=False,
+                _is_xpu=False,
+                _is_npu=False,
+                platforms=SimpleNamespace(
+                    current_platform=SimpleNamespace(
+                        is_cuda_alike=lambda: True,
+                        device_type="cuda",
+                    )
+                ),
+            ),
             patch(f"{BASE}.BaseImageProcessor", ImageProcessor),
             patch(f"{BASE}.torch.cuda.device", return_value=nullcontext()),
             patch(f"{BASE}.torch.cuda.MemPool", return_value="pool"),