[CI] Fix stale ServerArgs fake in chunked-SGMV LoRA test (#38315)

test/registered/kernels/ops/gemm/test_chunked_sgmv_cuda_graph.py

@@ -34,6 +34,7 @@ from sglang.kernels.ops.gemm.kv_b_lora_absorbed import (
 from sglang.srt.lora.backend.chunked_backend import ChunkedSgmvLoRABackend
 from sglang.srt.lora.utils import LoRABatchInfo
 from sglang.srt.model_executor.forward_batch_info import ForwardMode
+from sglang.srt.runtime_context import get_context
 from sglang.test.ci.ci_register import register_cuda_ci
 
 register_cuda_ci(est_time=60, stage="base-b-kernel-unit", runner_config="1-gpu-large")
@@ -186,13 +187,15 @@ def test_prepare_batch_neutralizes_static_tail_segments():
             self.batch_size = batch_size
             self.forward_mode = ForwardMode.DECODE
 
-    server_args = type("ServerArgs", (), {"max_lora_chunk_size": 16})
+    with get_context().override_server_args(max_lora_chunk_size=16) as server_args:
         backend = ChunkedSgmvLoRABackend(
             max_loras_per_batch=NUM_LORAS,
             device=torch.device("cuda"),
             server_args=server_args,
         )
-    backend.init_cuda_graph_batch_info(max_bs_in_cuda_graph=BS, num_tokens_per_req=1)
+        backend.init_cuda_graph_batch_info(
+            max_bs_in_cuda_graph=BS, num_tokens_per_req=1
+        )
         lora_ranks = [MAX_RANK] * NUM_LORAS
         scalings = [1.0] * NUM_LORAS