[CI] Reorganize stage-b 1-GPU tests for 5090 compatibility (#16826)

test/registered/kernels/test_fp4_moe.py

@@ -4,10 +4,6 @@ from typing import Callable
 import pytest
 import torch
 from flashinfer import fp4_quantize, scaled_fp4_grouped_quantize
-
-from sglang.test.ci.ci_register import register_cuda_ci
-
-register_cuda_ci(est_time=300, suite="nightly-4-gpu-b200", nightly=True)
 from flashinfer.fused_moe import cutlass_fused_moe as flashinfer_cutlass_fused_moe
 from sgl_kernel import scaled_fp4_quant, silu_and_mul
 from torch.nn import functional as F
@@ -15,6 +11,9 @@ from torch.nn import functional as F
 from sglang.srt.layers.moe.cutlass_moe import cutlass_moe_fp4
 from sglang.srt.layers.moe.cutlass_moe_params import CutlassMoEParams, CutlassMoEType
 from sglang.srt.layers.moe.topk import TopKConfig, select_experts
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=300, suite="nightly-4-gpu-b200", nightly=True)
 
 if torch.cuda.get_device_capability() < (10, 0):
     pytest.skip(

test/registered/kernels/test_nsa_indexer.py

@@ -4,11 +4,8 @@ from unittest.mock import MagicMock, patch
 
 import torch
 
-from sglang.test.ci.ci_register import register_cuda_ci
-
-register_cuda_ci(est_time=2, suite="nightly-1-gpu", nightly=True)
-
 from sglang.srt.layers import dp_attention as _dp_attn
+from sglang.test.ci.ci_register import register_cuda_ci
 
 # Patch DP-attention globals before importing backends
 _dp_attn.get_attention_tp_size = lambda: 1  # TP size = 1 for unit test
@@ -27,6 +24,8 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMo
 from sglang.srt.server_args import ServerArgs, set_global_server_args_for_scheduler
 from sglang.test.test_utils import CustomTestCase
 
+register_cuda_ci(est_time=2, suite="nightly-1-gpu", nightly=True)
+
 # Global configuration for all indexer tests
 DEFAULT_CONFIG = {
     "device": "cuda",