Piecewise Cuda Graph set default (#16331)

test/registered/models/test_generation_models.py

@@ -66,7 +66,7 @@ CI_MODELS = [
 # the complete set of models to test sglang's generation model
 ALL_MODELS = [
     *CI_MODELS,
-    ModelCase("Qwen/Qwen2-1.5B"),
+    ModelCase("Qwen/Qwen2-1.5B", decode_tolerance=7e-2),
     ModelCase("Qwen/Qwen2.5-14B-Instruct"),
     ModelCase("HuggingFaceTB/SmolLM-135M-Instruct", skip_long_prompt=True),
     ModelCase("allenai/OLMo-1B-0724-hf", decode_tolerance=8e-2, skip_long_prompt=True),

test/registered/models/test_kimi_linear_models_pcg.py

@@ -41,7 +41,6 @@ class TestKimiLinearPiecewiseCudaGraph(CustomTestCase):
                 "--tp",
                 "2",
                 "--trust-remote-code",
-                "--enable-piecewise-cuda-graph",
             ],
         )
 

test/registered/models/test_qwen3_next_models_pcg.py

@@ -40,9 +40,6 @@ class TestQwen3NextPiecewiseCudaGraph(CustomTestCase):
             other_args=[
                 "--tp",
                 "4",
-                "--enable-piecewise-cuda-graph",
-                "--piecewise-cuda-graph-compiler",
-                "eager",
             ],
         )