[AMD] Pin mem_fraction_static for the piecewise CUDA graph 1-GPU test on MI300 (#32862)

test/registered/cuda_graph/piecewise/test_piecewise_cuda_graph_support_1_gpu.py

@@ -13,6 +13,7 @@ from sglang.test.test_utils import (
     DEFAULT_URL_FOR_TEST,
     CustomTestCase,
     SimpleNamespace,
+    is_in_amd_ci,
     popen_launch_server,
 )
 
@@ -21,6 +22,13 @@ register_cuda_ci(est_time=180, suite="nightly-1-gpu", nightly=True)
 register_amd_ci(est_time=180, suite="stage-b-test-1-gpu-large-amd")
 
 
+# The 192GB mi300x runners have less headroom than the 256GB mi325x ones they
+# replaced: the auto-derived fraction left too little room for the ViT
+# activations plus the piecewise graph private pools, and the server died under
+# the 1024-thread gsm8k load.
+AMD_MEM_FRACTION_STATIC = 0.6
+
+
 class TestPiecewiseCudaGraphQwen25VL(CustomTestCase):
     """Test piecewise CUDA graph with Qwen2.5-VL-7B-Instruct model"""
 
@@ -28,14 +36,17 @@ class TestPiecewiseCudaGraphQwen25VL(CustomTestCase):
     def setUpClass(cls):
         cls.model = "Qwen/Qwen2.5-VL-7B-Instruct"
         cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--cuda-graph-backend-prefill=tc_piecewise",
+            "--disable-radix-cache",
+        ]
+        if is_in_amd_ci():
+            other_args += ["--mem-fraction-static", str(AMD_MEM_FRACTION_STATIC)]
         cls.process = popen_launch_server(
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-            other_args=[
+            other_args=other_args,
-                "--cuda-graph-backend-prefill=tc_piecewise",
-                "--disable-radix-cache",
-            ],
         )
 
     @classmethod
@@ -64,12 +75,16 @@ class TestPiecewiseCudaGraphQwen25VLEmbedding(CustomTestCase):
         model_path = "Qwen/Qwen2.5-VL-3B-Instruct"
         chat_template = get_chat_template_by_model_path(model_path)
         text = f"{chat_template.image_token}What is in this picture? Answer: "
+        extra_args = (
+            {"mem_fraction_static": AMD_MEM_FRACTION_STATIC} if is_in_amd_ci() else {}
+        )
 
         engine = Engine(
             model_path=model_path,
             enable_multimodal=True,
             is_embedding=True,
             cuda_graph_backend_prefill="tc_piecewise",
+            **extra_args,
         )
         out = engine.encode([text], image_data=[DEFAULT_IMAGE_URL])[0]["embedding"]
         engine.shutdown()
@@ -80,6 +95,7 @@ class TestPiecewiseCudaGraphQwen25VLEmbedding(CustomTestCase):
             enable_multimodal=True,
             is_embedding=True,
             cuda_graph_backend_prefill="disabled",
+            **extra_args,
         )
         out_without_pcg = engine.encode([text], image_data=[DEFAULT_IMAGE_URL])[0][
             "embedding"