[XPU] Enable XPU graph support (decode full-graph + prefill tc_piecewise) (#29053)

test/registered/xpu/test_deepseek_ocr.py

@@ -38,6 +38,7 @@ class TestDeepSeekOCR(CustomTestCase):
             "xpu",
             "--attention-backend",
             "intel_xpu",
+            "--disable-decode-cuda-graph",
         ]
         os.environ["SGLANG_USE_SGL_XPU"] = "1"
         cls.process = popen_launch_server(

test/registered/xpu/test_deepseek_ocr_triton.py

@@ -41,6 +41,7 @@ class TestDeepSeekOCRTriton(TestDeepSeekOCR):
             "xpu",
             "--attention-backend",
             "intel_xpu",
+            "--disable-decode-cuda-graph",
         ]
         os.environ["SGLANG_USE_SGL_XPU"] = "0"
         cls.process = popen_launch_server(

test/registered/xpu/test_encoder_attention_backend.py

@@ -39,6 +39,7 @@ class TestEncoderAttention(CustomTestCase):
             "xpu",
             "--mm-attention-backend",
             "xpu_attn",
+            "--disable-decode-cuda-graph",
         ]
         os.environ["SGLANG_USE_SGL_XPU"] = "1"
         cls.process = popen_launch_server(
@@ -127,6 +128,7 @@ class TestEncoderAttention_Triton(TestEncoderAttention):
             "xpu",
             "--mm-attention-backend",
             "triton_attn",
+            "--disable-decode-cuda-graph",
         ]
         os.environ["SGLANG_USE_SGL_XPU"] = "0"
         cls.process = popen_launch_server(

test/registered/xpu/test_gemma_4_e2b.py

@@ -53,6 +53,7 @@ XPU_SERVER_ARGS = [
     "intel_xpu",
     "--model-impl",
     "sglang",
+    "--disable-decode-cuda-graph",
 ]
 
 # Standard sglang e2e Q&A prompt (see test_openai_server.py::run_chat_completion).

test/registered/xpu/test_intel_xpu_backend.py

@@ -34,6 +34,7 @@ def intel_xpu_benchmark(
                 "1",
                 "--device",
                 "xpu",
+                "--disable-decode-cuda-graph",
             ]
             ci_args = ["--input", "64", "--output", "4"] if is_in_ci() else []
             full_args = common_args + ci_args + (extra_args or [])

test/registered/xpu/test_triton_attention_backend.py

@@ -31,6 +31,7 @@ def triton_attention_benchmark(extra_args=None, mem_fraction_static="0.84"):
                 "2050",
                 "--attention-backend",
                 "triton",
+                "--disable-decode-cuda-graph",
             ]
             full_args = common_args + (extra_args or [])
 

test/registered/xpu/test_xpu_basic.py

@@ -33,6 +33,7 @@ class TestXPUBasic(CustomTestCase):
             "0.6",
             "--batch-size",
             "1",
+            "--disable-decode-cuda-graph",
         ]
         if is_in_ci():
             args += ["--input", "64", "--output", "4"]

test/registered/xpu/test_xpu_embedding.py

@@ -34,7 +34,12 @@ class TestXPUEmbedding(CustomTestCase):
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-            other_args=["--is-embedding", "--device", "xpu"],
+            other_args=[
+                "--is-embedding",
+                "--device",
+                "xpu",
+                "--disable-decode-cuda-graph",
+            ],
         )
         cls.openai_url = cls.base_url + "/v1"
 

test/registered/xpu/test_xpu_graph.py

@@ -0,0 +1,71 @@
+"""
+XPU graph tests: verifies decode full-graph and prefill tc_piecewise graph
+on Intel XPU produce valid outputs.
+
+  - TestXPUGraph : decode full-graph and prefill tc_piecewise graph enabled
+    together in a single bench_one_batch invocation.
+
+Usage:
+    python3 -m unittest test_xpu_graph.TestXPUGraph
+"""
+
+import unittest
+
+from sglang.test.ci.ci_register import register_xpu_ci
+from sglang.test.test_utils import (
+    DEFAULT_SMALL_MODEL_NAME_FOR_TEST_QWEN,
+    CustomTestCase,
+    is_in_ci,
+    run_bench_one_batch,
+)
+
+register_xpu_ci(est_time=600, suite="stage-b-test-1-gpu-xpu")
+
+_COMMON_ARGS = [
+    "--device",
+    "xpu",
+    "--attention-backend",
+    "triton",
+    "--disable-radix-cache",
+    "--mem-fraction-static",
+    "0.6",
+    "--batch-size",
+    "1",
+]
+
+_CI_IO_ARGS = ["--input", "64", "--output", "4"]
+_FULL_IO_ARGS = ["--input", "128", "--output", "16"]
+
+
+class TestXPUGraph(CustomTestCase):
+    """Decode full-graph + prefill tc_piecewise together."""
+
+    def test_full_graph_runs(self):
+        args = [
+            *_COMMON_ARGS,
+            "--cuda-graph-config",
+            '{"decode":{"backend":"full"},"prefill":{"backend":"tc_piecewise","tc_compiler":"eager"}}',
+            "--cuda-graph-bs-prefill",
+            "64",
+            "128",
+        ]
+        if is_in_ci():
+            args += _CI_IO_ARGS
+        else:
+            args += _FULL_IO_ARGS
+
+        prefill_latency, decode_throughput, _ = run_bench_one_batch(
+            DEFAULT_SMALL_MODEL_NAME_FOR_TEST_QWEN, args
+        )
+        self.assertGreater(
+            prefill_latency,
+            0,
+            "prefill latency must be > 0 with tc_piecewise XPU graph",
+        )
+        self.assertGreater(
+            decode_throughput, 0, "decode throughput must be > 0 with full XPU graph"
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/xpu/test_xpu_serving_features.py

@@ -43,7 +43,7 @@ class TestXPUServingFeatures(CustomTestCase):
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-            other_args=["--device", "xpu"],
+            other_args=["--device", "xpu", "--disable-decode-cuda-graph"],
         )
         cls.openai_url = cls.base_url + "/v1"