[XPU] Remove redundant xpu graph backend and make xpu graph opt-in by default (#29911)

test/registered/spec/eagle/test_spec_eagle_parity.py

@@ -42,6 +42,10 @@ class TestEagle3ParityXPU(SpecParityKit, _Eagle3ParityBase):
 
     disable_overlap = False
     attention_backend = "triton"
+    # Decode full-graph was active by default when this test was added
+    # (via XPUCudaGraphBackend). Opt in explicitly now that it is disabled
+    # by default so the coverage is preserved.
+    extra_args = ("--cuda-graph-config", '{"decode":{"backend":"full"}}')
 
 
 if __name__ == "__main__":

test/registered/xpu/test_deepseek_ocr.py

@@ -38,7 +38,6 @@ 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

@@ -41,7 +41,6 @@ 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

@@ -39,7 +39,6 @@ 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(
@@ -128,7 +127,6 @@ 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

@@ -53,7 +53,6 @@ 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

@@ -34,7 +34,6 @@ 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

@@ -31,7 +31,6 @@ 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

@@ -33,7 +33,6 @@ 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

@@ -38,7 +38,6 @@ class TestXPUEmbedding(CustomTestCase):
                 "--is-embedding",
                 "--device",
                 "xpu",
-                "--disable-decode-cuda-graph",
             ],
         )
         cls.openai_url = cls.base_url + "/v1"

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", "--disable-decode-cuda-graph"],
+            other_args=["--device", "xpu"],
         )
         cls.openai_url = cls.base_url + "/v1"