fix scheduler for non-cuda devices and disable piecewise cuda graph f… (#19992)

Co-authored-by: Kangyan-Zhou <zky314343421@gmail.com>

.github/workflows/pr-test-xpu.yml

@@ -114,15 +114,13 @@ jobs:
           docker exec "$cid" /home/sdp/miniforge3/envs/py3.10/bin/python3 -m pip install pytest expecttest ray huggingface_hub
           docker exec "$cid" /home/sdp/miniforge3/envs/py3.10/bin/python3 -m pip uninstall -y flashinfer-python
           docker exec "$cid" /bin/bash -c '/home/sdp/miniforge3/envs/py3.10/bin/hf auth login --token ${HF_TOKEN} '
-          docker exec -u root "$cid" /bin/bash -c "ln -sf /home/sdp/miniforge3/envs/py3.10/bin/python3 /usr/bin/python3"
+
 
       - name: Run E2E Bfloat16 tests
         timeout-minutes: 20
         run: |
           cid="${{ steps.start_container.outputs.container_id }}"
-          docker exec -w /home/sdp/sglang/ "$cid" \
+          docker exec "$cid" bash -c "source /home/sdp/miniforge3/bin/activate && conda activate py3.10 && cd /home/sdp/sglang/test/srt && python3 run_suite.py --suite per-commit-xpu"
-            bash -c "LD_LIBRARY_PATH=/home/sdp/miniforge3/envs/py3.10/lib:$LD_LIBRARY_PATH && cd ./test/srt && python3 run_suite.py --suite per-commit-xpu"
-
       - name: Cleanup container
         if: always()
         run: |

python/sglang/srt/server_args.py

@@ -59,6 +59,7 @@ from sglang.srt.utils.common import (
     is_sm100_supported,
     is_sm120_supported,
     is_triton_kernels_available,
+    is_xpu,
     json_list_type,
     nullable_str,
     parse_connector_type,
@@ -1067,8 +1068,8 @@ class ServerArgs:
         # 5. Pipeline parallelism
         if self.pp_size > 1:
             self.disable_piecewise_cuda_graph = True
-        # 6. Non-CUDA hardware (AMD, NPU, CPU, MPS, MUSA, etc.)
+        # 6. Non-CUDA hardware (AMD, NPU, CPU, etc.)
-        if is_hip() or is_npu() or is_cpu() or is_mps() or is_musa():
+        if is_hip() or is_npu() or is_cpu() or is_mps() or is_musa() or is_xpu():
             self.disable_piecewise_cuda_graph = True
         # 7. MoE A2A backend
         if self.moe_a2a_backend != "none":

test/srt/run_suite.py

@@ -75,7 +75,7 @@ suite_xeon = {
 suite_xpu = {
     "per-commit-xpu": [
         TestFile("xpu/test_deepseek_ocr.py"),
-        TestFile("xpu/test_internvl.py"),
+        # TestFile("xpu/test_internvl.py"),
         TestFile("xpu/test_intel_xpu_backend.py"),
     ],
 }

test/srt/xpu/test_internvl.py

@@ -1,147 +0,0 @@
-"""
-XPU tests for InternVL models (InternVL2.5-2B, InternVL3.5-2B).
-
-Uses the same structure as test_vision_openai_server_a.py: OpenAI /v1 chat API
-and ImageOpenAITestMixin. An XPU-specific base injects --device xpu and
---attention-backend intel_xpu.
-
-Usage (pick module path to match your cwd):
-
-  From test/srt/xpu:
-    python3 -m unittest test_internvl.TestInternVL25Server.test_single_image_chat_completion
-    python3 -m unittest test_internvl
-
-  From test/srt:
-    python3 -m unittest xpu.test_internvl.TestInternVL25Server.test_single_image_chat_completion
-    python3 -m unittest xpu.test_internvl
-
-  From repo root:
-    python3 -m unittest test.srt.xpu.test_internvl.TestInternVL25Server.test_single_image_chat_completion
-    python3 -m unittest test.srt.xpu.test_internvl
-"""
-
-import os
-import unittest
-
-from sglang.test.vlm_utils import (
-    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-    DEFAULT_URL_FOR_TEST,
-    ImageOpenAITestMixin,
-    TestOpenAIMLLMServerBase,
-    kill_process_tree,
-    popen_launch_server,
-)
-
-# XPU args injected into server launch for all InternVL XPU tests
-XPU_ARGS = [
-    "--device",
-    "xpu",
-    "--attention-backend",
-    "intel_xpu",
-]
-
-
-# Longer launch timeout for InternVL3.5 (can be slow to start on XPU)
-INTERNVL35_LAUNCH_TIMEOUT = 900
-
-
-class InternVLXPUServerBase(TestOpenAIMLLMServerBase):
-    """Base for InternVL tests on XPU. Injects XPU args and sets SGLANG_USE_SGL_XPU."""
-
-    use_sgl_xpu = True  # subclasses override for Triton backend
-    launch_timeout = None  # subclasses can set to override (seconds)
-
-    @classmethod
-    def setUpClass(cls):
-        cls.base_url = DEFAULT_URL_FOR_TEST
-        cls.api_key = "sk-123456"
-
-        os.environ["SGLANG_USE_SGL_XPU"] = "1" if cls.use_sgl_xpu else "0"
-
-        other_args = list(XPU_ARGS) + list(cls.extra_args)
-        if cls.trust_remote_code:
-            other_args.extend(cls.fixed_args)
-        else:
-            other_args.extend(
-                arg for arg in cls.fixed_args if arg != "--trust-remote-code"
-            )
-
-        timeout = (
-            cls.launch_timeout
-            if cls.launch_timeout is not None
-            else DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
-        )
-        cls.process = popen_launch_server(
-            cls.model,
-            cls.base_url,
-            timeout=timeout,
-            api_key=cls.api_key,
-            other_args=other_args,
-        )
-        cls.base_url += "/v1"
-
-    @classmethod
-    def tearDownClass(cls):
-        kill_process_tree(cls.process.pid)
-
-
-class TestInternVL25Server(ImageOpenAITestMixin, InternVLXPUServerBase):
-    """InternVL2.5-2B on XPU with SGL XPU backend."""
-
-    model = "OpenGVLab/InternVL2_5-2B"
-    use_sgl_xpu = True
-    extra_args = [
-        "--cuda-graph-max-bs=4",
-    ]
-
-    def test_video_images_chat_completion(self):
-        # Video test uses 10 frames and exceeds max_prefill_tokens (23124 > 16384).
-        pass
-
-
-class TestInternVL25TritonServer(ImageOpenAITestMixin, InternVLXPUServerBase):
-    """InternVL2.5-2B on XPU with Triton (non-SGL) backend."""
-
-    model = "OpenGVLab/InternVL2_5-2B"
-    use_sgl_xpu = False
-    extra_args = [
-        "--cuda-graph-max-bs=4",
-    ]
-
-    def test_video_images_chat_completion(self):
-        # Video test exceeds max_prefill_tokens on XPU with default limits.
-        pass
-
-
-class TestInternVL35_2BServer(ImageOpenAITestMixin, InternVLXPUServerBase):
-    """InternVL3.5-2B on XPU with SGL XPU backend."""
-
-    model = "OpenGVLab/InternVL3_5-2B"
-    use_sgl_xpu = True
-    launch_timeout = INTERNVL35_LAUNCH_TIMEOUT
-    extra_args = [
-        "--cuda-graph-max-bs=4",
-    ]
-
-    def test_video_images_chat_completion(self):
-        # Video test exceeds max_prefill_tokens (23202 > 14588) on InternVL3.5.
-        pass
-
-
-class TestInternVL35_2BTritonServer(ImageOpenAITestMixin, InternVLXPUServerBase):
-    """InternVL3.5-2B on XPU with Triton (non-SGL) backend."""
-
-    model = "OpenGVLab/InternVL3_5-2B"
-    use_sgl_xpu = False
-    launch_timeout = INTERNVL35_LAUNCH_TIMEOUT
-    extra_args = [
-        "--cuda-graph-max-bs=4",
-    ]
-
-    def test_video_images_chat_completion(self):
-        # Video test exceeds max_prefill_tokens on InternVL3.5.
-        pass
-
-
-if __name__ == "__main__":
-    unittest.main()