diff --git a/test/registered/4-gpu-models/test_gpt_oss_4gpu.py b/test/registered/4-gpu-models/test_gpt_oss_4gpu.py
index 5e4ceef2a..dfbfc4fb1 100644
--- a/test/registered/4-gpu-models/test_gpt_oss_4gpu.py
+++ b/test/registered/4-gpu-models/test_gpt_oss_4gpu.py
@@ -3,8 +3,8 @@ import unittest
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.gpt_oss_common import BaseTestGptOss
 
-register_cuda_ci(est_time=300, suite="stage-c-test-4-gpu-h100")
-register_cuda_ci(est_time=300, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=328, suite="stage-c-test-4-gpu-h100")
+register_cuda_ci(est_time=312, suite="stage-c-test-4-gpu-b200")
 
 
 class TestGptOss4Gpu(BaseTestGptOss):
diff --git a/test/registered/4-gpu-models/test_nvidia_nemotron_3_super_nvfp4.py b/test/registered/4-gpu-models/test_nvidia_nemotron_3_super_nvfp4.py
index f658b1b86..f1d68e69b 100644
--- a/test/registered/4-gpu-models/test_nvidia_nemotron_3_super_nvfp4.py
+++ b/test/registered/4-gpu-models/test_nvidia_nemotron_3_super_nvfp4.py
@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=290, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=294, suite="stage-c-test-4-gpu-b200")
 
 NEMOTRON_3_SUPER_NVFP4_MODEL = "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4"
 
diff --git a/test/registered/4-gpu-models/test_qwen35_hicache.py b/test/registered/4-gpu-models/test_qwen35_hicache.py
index ac001ff32..66b6cd9f3 100644
--- a/test/registered/4-gpu-models/test_qwen35_hicache.py
+++ b/test/registered/4-gpu-models/test_qwen35_hicache.py
@@ -18,7 +18,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=620, suite="stage-c-test-4-gpu-h100")
+register_cuda_ci(est_time=540, suite="stage-c-test-4-gpu-h100")
 
 QWEN35_27B_MODEL = "Qwen/Qwen3.5-27B"
 ACC_THRESHOLDS = {QWEN35_27B_MODEL: {"gsm8k": 0.8}}
diff --git a/test/registered/4-gpu-models/test_qwen35_models.py b/test/registered/4-gpu-models/test_qwen35_models.py
index 828857b55..667c62abd 100644
--- a/test/registered/4-gpu-models/test_qwen35_models.py
+++ b/test/registered/4-gpu-models/test_qwen35_models.py
@@ -21,7 +21,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=790, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=768, suite="stage-c-test-4-gpu-b200")
 
 QWEN35_FP4_MODEL = "nvidia/Qwen3.5-397B-A17B-NVFP4"
 ACC_THRESHOLDS = {QWEN35_FP4_MODEL: {"gsm8k": 0.95}}
diff --git a/test/registered/4-gpu-models/test_qwen3_30b.py b/test/registered/4-gpu-models/test_qwen3_30b.py
index 1e7ebbe33..079dbc985 100644
--- a/test/registered/4-gpu-models/test_qwen3_30b.py
+++ b/test/registered/4-gpu-models/test_qwen3_30b.py
@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=150, suite="stage-c-test-4-gpu-h100")
+register_cuda_ci(est_time=146, suite="stage-c-test-4-gpu-h100")
 
 QWEN3_30B_MODEL_PATH = "Qwen/Qwen3-30B-A3B-FP8"
 
diff --git a/test/registered/4-gpu-models/test_qwen3_next_models.py b/test/registered/4-gpu-models/test_qwen3_next_models.py
index 48b711dd7..c9bc1839e 100644
--- a/test/registered/4-gpu-models/test_qwen3_next_models.py
+++ b/test/registered/4-gpu-models/test_qwen3_next_models.py
@@ -6,7 +6,7 @@ from sglang.test.kits.kl_divergence_kit import KLDivergenceMixin
 from sglang.test.kits.prefix_cache_branching_kit import PrefixCacheBranchingMixin
 from sglang.test.server_fixtures.default_fixture import DefaultServerBase
 
-register_cuda_ci(est_time=120, suite="stage-c-test-4-gpu-h100")
+register_cuda_ci(est_time=119, suite="stage-c-test-4-gpu-h100")
 
 QWEN3_NEXT_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct"
 
diff --git a/test/registered/4-gpu-models/test_qwen3_next_models_mtp.py b/test/registered/4-gpu-models/test_qwen3_next_models_mtp.py
index 3be1ddd0c..e7edfbee9 100644
--- a/test/registered/4-gpu-models/test_qwen3_next_models_mtp.py
+++ b/test/registered/4-gpu-models/test_qwen3_next_models_mtp.py
@@ -7,7 +7,7 @@ from sglang.test.kits.kl_divergence_kit import KLDivergenceMixin
 from sglang.test.kits.prefix_cache_branching_kit import PrefixCacheBranchingMixin
 from sglang.test.server_fixtures.default_fixture import DefaultServerBase
 
-register_cuda_ci(est_time=430, suite="stage-c-test-4-gpu-h100")
+register_cuda_ci(est_time=422, suite="stage-c-test-4-gpu-h100")
 
 QWEN3_NEXT_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct"
 
diff --git a/test/registered/8-gpu-models/test_deepseek_v32_indexcache.py b/test/registered/8-gpu-models/test_deepseek_v32_indexcache.py
index 4b769ec57..320e23028 100644
--- a/test/registered/8-gpu-models/test_deepseek_v32_indexcache.py
+++ b/test/registered/8-gpu-models/test_deepseek_v32_indexcache.py
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=360, suite="stage-c-test-8-gpu-h200")
+register_cuda_ci(est_time=354, suite="stage-c-test-8-gpu-h200")
 
 DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
 
diff --git a/test/registered/8-gpu-models/test_deepseek_v3_basic.py b/test/registered/8-gpu-models/test_deepseek_v3_basic.py
index acbde5475..c814ed127 100644
--- a/test/registered/8-gpu-models/test_deepseek_v3_basic.py
+++ b/test/registered/8-gpu-models/test_deepseek_v3_basic.py
@@ -15,7 +15,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=275, suite="stage-c-test-8-gpu-h200")
+register_cuda_ci(est_time=320, suite="stage-c-test-8-gpu-h200")
 
 FULL_DEEPSEEK_V3_MODEL_PATH = "deepseek-ai/DeepSeek-V3-0324"
 
diff --git a/test/registered/8-gpu-models/test_deepseek_v3_mtp.py b/test/registered/8-gpu-models/test_deepseek_v3_mtp.py
index e90cd4fb0..805db513d 100644
--- a/test/registered/8-gpu-models/test_deepseek_v3_mtp.py
+++ b/test/registered/8-gpu-models/test_deepseek_v3_mtp.py
@@ -17,7 +17,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=240, suite="stage-c-test-8-gpu-h200")
+register_cuda_ci(est_time=198, suite="stage-c-test-8-gpu-h200")
 
 FULL_DEEPSEEK_V3_MODEL_PATH = "deepseek-ai/DeepSeek-V3-0324"
 
diff --git a/test/registered/8-gpu-models/test_dsa_models_basic.py b/test/registered/8-gpu-models/test_dsa_models_basic.py
index 3a7b405ea..cc6caffd2 100644
--- a/test/registered/8-gpu-models/test_dsa_models_basic.py
+++ b/test/registered/8-gpu-models/test_dsa_models_basic.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=870, suite="stage-c-test-8-gpu-h200")
+register_cuda_ci(est_time=966, suite="stage-c-test-8-gpu-h200")
 
 DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
 GLM5_MODEL_PATH = "zai-org/GLM-5-FP8"
diff --git a/test/registered/8-gpu-models/test_dsa_models_mtp.py b/test/registered/8-gpu-models/test_dsa_models_mtp.py
index 55457f9fd..9c0b2dcef 100644
--- a/test/registered/8-gpu-models/test_dsa_models_mtp.py
+++ b/test/registered/8-gpu-models/test_dsa_models_mtp.py
@@ -17,7 +17,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=880, suite="stage-c-test-8-gpu-h200")
+register_cuda_ci(est_time=910, suite="stage-c-test-8-gpu-h200")
 
 FULL_DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
 GLM5_MODEL_PATH = "zai-org/GLM-5-FP8"
diff --git a/test/registered/8-gpu-models/test_mimo_models.py b/test/registered/8-gpu-models/test_mimo_models.py
index aa9da353e..fe79139bd 100644
--- a/test/registered/8-gpu-models/test_mimo_models.py
+++ b/test/registered/8-gpu-models/test_mimo_models.py
@@ -5,7 +5,7 @@ from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
 from sglang.test.kits.spec_decoding_kit import SpecDecodingMixin
 from sglang.test.server_fixtures.default_fixture import DefaultServerBase
 
-register_cuda_ci(est_time=270, suite="stage-c-test-8-gpu-h200")
+register_cuda_ci(est_time=283, suite="stage-c-test-8-gpu-h200")
 
 
 class TestMiMoV2Flash(GSM8KMixin, SpecDecodingMixin, DefaultServerBase):
diff --git a/test/registered/8-gpu-models/test_minimax_m25_basic.py b/test/registered/8-gpu-models/test_minimax_m25_basic.py
index 61837d479..f99e4c3cd 100644
--- a/test/registered/8-gpu-models/test_minimax_m25_basic.py
+++ b/test/registered/8-gpu-models/test_minimax_m25_basic.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=300, suite="stage-c-test-8-gpu-h200")
+register_cuda_ci(est_time=330, suite="stage-c-test-8-gpu-h200")
 
 MINIMAX_M25_MODEL_PATH = "MiniMaxAI/MiniMax-M2.5"
 
diff --git a/test/registered/8-gpu-models/test_nvidia_nemotron_3_super_bf16.py b/test/registered/8-gpu-models/test_nvidia_nemotron_3_super_bf16.py
index 973a5019a..5d2200f3c 100644
--- a/test/registered/8-gpu-models/test_nvidia_nemotron_3_super_bf16.py
+++ b/test/registered/8-gpu-models/test_nvidia_nemotron_3_super_bf16.py
@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=300, suite="stage-c-test-8-gpu-h200")
+register_cuda_ci(est_time=319, suite="stage-c-test-8-gpu-h200")
 
 NEMOTRON_3_SUPER_BF16_MODEL = "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16"
 
diff --git a/test/registered/attention/test_chunk_gated_delta_rule.py b/test/registered/attention/test_chunk_gated_delta_rule.py
index 5daeb6a1b..f1ac1a40f 100644
--- a/test/registered/attention/test_chunk_gated_delta_rule.py
+++ b/test/registered/attention/test_chunk_gated_delta_rule.py
@@ -8,7 +8,7 @@ from sglang.srt.layers.attention.fla.fused_recurrent import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=5, suite="stage-b-test-1-gpu-large")
 
 
 @unittest.skipIf(not torch.cuda.is_available(), "Test requires CUDA")
diff --git a/test/registered/attention/test_create_kvindices.py b/test/registered/attention/test_create_kvindices.py
index 0b06e8222..c1bd28cac 100644
--- a/test/registered/attention/test_create_kvindices.py
+++ b/test/registered/attention/test_create_kvindices.py
@@ -9,7 +9,7 @@ from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 # Triton kernel unit test for KV indices creation
-register_cuda_ci(est_time=10, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=7, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=10, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/attention/test_fa3.py b/test/registered/attention/test_fa3.py
index d0c48cfcd..b941e134d 100644
--- a/test/registered/attention/test_fa3.py
+++ b/test/registered/attention/test_fa3.py
@@ -20,7 +20,7 @@ from sglang.test.test_utils import (
 
 # FlashAttention3 integration tests (requires SM 90+ / H100)
 # Multiple test classes: FA3, FA3+MLA, FA3+SpecDecode variants
-register_cuda_ci(est_time=390, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=386, suite="stage-b-test-1-gpu-large")
 
 GSM_DATASET_PATH = None
 
diff --git a/test/registered/attention/test_flash_attention_4.py b/test/registered/attention/test_flash_attention_4.py
index ce61ec252..cae354fe5 100644
--- a/test/registered/attention/test_flash_attention_4.py
+++ b/test/registered/attention/test_flash_attention_4.py
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
 )
 
 # FlashAttention4 integration test (requires SM 100+ / Blackwell B200)
-register_cuda_ci(est_time=200, suite="stage-b-test-4-gpu-b200")
+register_cuda_ci(est_time=259, suite="stage-b-test-4-gpu-b200")
 
 
 @unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")
diff --git a/test/registered/attention/test_hybrid_attn_backend.py b/test/registered/attention/test_hybrid_attn_backend.py
index cb41c0cbd..dce9ca3c6 100644
--- a/test/registered/attention/test_hybrid_attn_backend.py
+++ b/test/registered/attention/test_hybrid_attn_backend.py
@@ -20,7 +20,7 @@ from sglang.test.test_utils import (
 
 # Hybrid attention backend tests (FA3 prefill + FlashInfer decode, requires SM 90+ / H100)
 # Multiple test classes: base, MLA, TorchCompile, SpecDecode variants
-register_cuda_ci(est_time=350, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=342, suite="stage-b-test-1-gpu-large")
 
 GSM_DATASET_PATH = None
 
diff --git a/test/registered/attention/test_kda_kernels.py b/test/registered/attention/test_kda_kernels.py
index d7a971934..f3b18ad59 100644
--- a/test/registered/attention/test_kda_kernels.py
+++ b/test/registered/attention/test_kda_kernels.py
@@ -8,7 +8,7 @@ from sglang.srt.layers.attention.fla.fused_sigmoid_gating_recurrent import (
 from sglang.srt.layers.attention.fla.kda import fused_kda_gate, fused_recurrent_kda
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=30, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-large")
 
 
 @unittest.skipIf(not torch.cuda.is_available(), "Test requires CUDA")
diff --git a/test/registered/attention/test_local_attn.py b/test/registered/attention/test_local_attn.py
index 229260c19..f14be8cf7 100644
--- a/test/registered/attention/test_local_attn.py
+++ b/test/registered/attention/test_local_attn.py
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
 )
 
 # Local attention with FA3 (requires SM 90+ / H100, tp=4)
-register_cuda_ci(est_time=200, suite="stage-c-test-4-gpu-h100")
+register_cuda_ci(est_time=208, suite="stage-c-test-4-gpu-h100")
 
 
 @unittest.skipIf(get_device_sm() < 90, "Test requires CUDA SM 90 or higher")
diff --git a/test/registered/attention/test_normal_decode_set_metadata.py b/test/registered/attention/test_normal_decode_set_metadata.py
index 21e701521..6906f6cfa 100644
--- a/test/registered/attention/test_normal_decode_set_metadata.py
+++ b/test/registered/attention/test_normal_decode_set_metadata.py
@@ -21,7 +21,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 # Register this test for CUDA CI in stage-b (fast attention/kernel tests)
-register_cuda_ci(est_time=25, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=9, suite="stage-b-test-1-gpu-large")
 
 
 def reference_normal_decode_set_metadata(
diff --git a/test/registered/attention/test_torch_native_attention_backend.py b/test/registered/attention/test_torch_native_attention_backend.py
index 0d2c9f330..102a7816b 100644
--- a/test/registered/attention/test_torch_native_attention_backend.py
+++ b/test/registered/attention/test_torch_native_attention_backend.py
@@ -18,7 +18,7 @@ from sglang.test.test_utils import (
 )
 
 # Torch native attention backend integration test with MMLU eval
-register_cuda_ci(est_time=169, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=128, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=150, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/attention/test_triton_attention_backend.py b/test/registered/attention/test_triton_attention_backend.py
index 47900a20a..2034a89a6 100644
--- a/test/registered/attention/test_triton_attention_backend.py
+++ b/test/registered/attention/test_triton_attention_backend.py
@@ -20,7 +20,7 @@ from sglang.test.test_utils import (
 )
 
 # Triton attention backend integration test with latency benchmark and MMLU eval
-register_cuda_ci(est_time=200, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=164, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=1400, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/attention/test_triton_attention_kernels.py b/test/registered/attention/test_triton_attention_kernels.py
index 89bc81ee0..48b74cb72 100644
--- a/test/registered/attention/test_triton_attention_kernels.py
+++ b/test/registered/attention/test_triton_attention_kernels.py
@@ -23,7 +23,7 @@ from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase, is_in_amd_ci
 
 # Triton attention kernel unit tests (decode, extend, prefill)
-register_cuda_ci(est_time=30, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=16, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=30, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/attention/test_triton_sliding_window.py b/test/registered/attention/test_triton_sliding_window.py
index dbc4ba3d5..faff86108 100644
--- a/test/registered/attention/test_triton_sliding_window.py
+++ b/test/registered/attention/test_triton_sliding_window.py
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
 )
 
 # Sliding window attention with Triton backend (Gemma-3 model)
-register_cuda_ci(est_time=100, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=95, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=200, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/backends/test_torch_compile.py b/test/registered/backends/test_torch_compile.py
index 6afce40e8..3f5f1ca91 100644
--- a/test/registered/backends/test_torch_compile.py
+++ b/test/registered/backends/test_torch_compile.py
@@ -15,7 +15,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=144, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=115, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=1100, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/bench_fn/test_benchmark_datasets_api.py b/test/registered/bench_fn/test_benchmark_datasets_api.py
index 82c8573d7..d807b5033 100644
--- a/test/registered/bench_fn/test_benchmark_datasets_api.py
+++ b/test/registered/bench_fn/test_benchmark_datasets_api.py
@@ -26,7 +26,7 @@ from sglang.benchmark.datasets.random import sample_random_requests
 from sglang.benchmark.datasets.sharegpt import sample_sharegpt_requests
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=6, suite="stage-a-test-cpu")
 
 
 class _DummyTokenTensor:
diff --git a/test/registered/constrained_decoding/test_constrained_decoding.py b/test/registered/constrained_decoding/test_constrained_decoding.py
index b705b6f99..9e3cf9213 100644
--- a/test/registered/constrained_decoding/test_constrained_decoding.py
+++ b/test/registered/constrained_decoding/test_constrained_decoding.py
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=111, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=110, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=179, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/core/test_deterministic.py b/test/registered/core/test_deterministic.py
index 956504929..7b269a93f 100644
--- a/test/registered/core/test_deterministic.py
+++ b/test/registered/core/test_deterministic.py
@@ -16,7 +16,7 @@ from sglang.test.test_deterministic_utils import (
 )
 from sglang.test.test_utils import is_in_amd_ci
 
-register_cuda_ci(est_time=200, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=194, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=278, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/core/test_gpt_oss_1gpu.py b/test/registered/core/test_gpt_oss_1gpu.py
index c208e8077..692af9070 100644
--- a/test/registered/core/test_gpt_oss_1gpu.py
+++ b/test/registered/core/test_gpt_oss_1gpu.py
@@ -3,7 +3,7 @@ import unittest
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.gpt_oss_common import BaseTestGptOss
 
-register_cuda_ci(est_time=420, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=372, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=750, suite="stage-b-test-1-gpu-small-amd-mi35x")
 
 
diff --git a/test/registered/core/test_gpt_oss_sm120.py b/test/registered/core/test_gpt_oss_sm120.py
index dcbe427e7..18ad85478 100644
--- a/test/registered/core/test_gpt_oss_sm120.py
+++ b/test/registered/core/test_gpt_oss_sm120.py
@@ -5,7 +5,7 @@ import torch
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.gpt_oss_common import BaseTestGptOss
 
-register_cuda_ci(est_time=330, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=329, suite="stage-b-test-1-gpu-small")
 
 
 @unittest.skipIf(not torch.cuda.is_available(), "CUDA is not available")
diff --git a/test/registered/core/test_hidden_states.py b/test/registered/core/test_hidden_states.py
index 20c4db95e..1e8d8a764 100644
--- a/test/registered/core/test_hidden_states.py
+++ b/test/registered/core/test_hidden_states.py
@@ -8,7 +8,7 @@ from sglang.srt.utils import get_device, is_hip
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import DEFAULT_SMALL_MODEL_NAME_FOR_TEST, CustomTestCase
 
-register_cuda_ci(est_time=55, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=47, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=55, suite="stage-b-test-1-gpu-small-amd")
 
 _is_hip = is_hip()
diff --git a/test/registered/core/test_page_size.py b/test/registered/core/test_page_size.py
index 1cf8cbcb6..bfd8d116c 100644
--- a/test/registered/core/test_page_size.py
+++ b/test/registered/core/test_page_size.py
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=60, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=77, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=60, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/core/test_request_queue_validation.py b/test/registered/core/test_request_queue_validation.py
index 36727c4de..d6293ae2f 100644
--- a/test/registered/core/test_request_queue_validation.py
+++ b/test/registered/core/test_request_queue_validation.py
@@ -17,7 +17,7 @@ from sglang.test.test_utils import (
     send_generate_requests,
 )
 
-register_cuda_ci(est_time=47, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=50, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=70, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/core/test_srt_endpoint.py b/test/registered/core/test_srt_endpoint.py
index 22e3b468f..a639d7861 100644
--- a/test/registered/core/test_srt_endpoint.py
+++ b/test/registered/core/test_srt_endpoint.py
@@ -27,7 +27,7 @@ from sglang.test.test_utils import (
     run_logprob_check,
 )
 
-register_cuda_ci(est_time=127, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=132, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=130, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/core/test_srt_engine.py b/test/registered/core/test_srt_engine.py
index ab70eaaf1..f6610ddf8 100644
--- a/test/registered/core/test_srt_engine.py
+++ b/test/registered/core/test_srt_engine.py
@@ -22,7 +22,7 @@ from sglang.test.test_utils import (
     CustomTestCase,
 )
 
-register_cuda_ci(est_time=252, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=311, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=261, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/cp/test_deepseek_v32_cp_single_node.py b/test/registered/cp/test_deepseek_v32_cp_single_node.py
index 55595eff5..badcb2caa 100644
--- a/test/registered/cp/test_deepseek_v32_cp_single_node.py
+++ b/test/registered/cp/test_deepseek_v32_cp_single_node.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=640, suite="stage-c-test-deepep-8-gpu-h200")
+register_cuda_ci(est_time=614, suite="stage-c-test-deepep-8-gpu-h200")
 DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
 
 
diff --git a/test/registered/disaggregation/test_disaggregation_basic.py b/test/registered/disaggregation/test_disaggregation_basic.py
index a598ae438..fd63195b3 100644
--- a/test/registered/disaggregation/test_disaggregation_basic.py
+++ b/test/registered/disaggregation/test_disaggregation_basic.py
@@ -20,7 +20,7 @@ from sglang.test.test_utils import (
     DEFAULT_TARGET_MODEL_EAGLE3,
 )
 
-register_cuda_ci(est_time=400, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=394, suite="stage-b-test-2-gpu-large")
 
 
 class TestDisaggregationAccuracy(PDDisaggregationServerBase):
diff --git a/test/registered/disaggregation/test_specv2_kvcache_offloading.py b/test/registered/disaggregation/test_specv2_kvcache_offloading.py
index 5c403ba2e..639d5b44f 100644
--- a/test/registered/disaggregation/test_specv2_kvcache_offloading.py
+++ b/test/registered/disaggregation/test_specv2_kvcache_offloading.py
@@ -17,7 +17,7 @@ from sglang.srt.disaggregation.decode_kvcache_offload_manager import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=10, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-small")
 
 
 def _make_mock_req(
diff --git a/test/registered/distributed/test_data_parallelism.py b/test/registered/distributed/test_data_parallelism.py
index bb3eb29c7..a7d49ac19 100644
--- a/test/registered/distributed/test_data_parallelism.py
+++ b/test/registered/distributed/test_data_parallelism.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=73, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=78, suite="stage-b-test-2-gpu-large")
 register_amd_ci(est_time=73, suite="stage-b-test-2-gpu-large-amd")
 
 
diff --git a/test/registered/distributed/test_disaggregation_different_tp.py b/test/registered/distributed/test_disaggregation_different_tp.py
index 7ea78bfbf..3fd1a9504 100644
--- a/test/registered/distributed/test_disaggregation_different_tp.py
+++ b/test/registered/distributed/test_disaggregation_different_tp.py
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=330, suite="stage-c-test-8-gpu-h20")
+register_cuda_ci(est_time=331, suite="stage-c-test-8-gpu-h20")
 
 
 class TestDisaggregationMooncakePrefillLargerTP(PDDisaggregationServerBase):
diff --git a/test/registered/distributed/test_disaggregation_dp_attention.py b/test/registered/distributed/test_disaggregation_dp_attention.py
index b76cf1b18..3eab31142 100644
--- a/test/registered/distributed/test_disaggregation_dp_attention.py
+++ b/test/registered/distributed/test_disaggregation_dp_attention.py
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=400, suite="stage-c-test-8-gpu-h20")
+register_cuda_ci(est_time=408, suite="stage-c-test-8-gpu-h20")
 
 
 class TestDisaggregationDPAttention(PDDisaggregationServerBase):
diff --git a/test/registered/distributed/test_disaggregation_hybrid_attention.py b/test/registered/distributed/test_disaggregation_hybrid_attention.py
index edbb177eb..87ef16974 100644
--- a/test/registered/distributed/test_disaggregation_hybrid_attention.py
+++ b/test/registered/distributed/test_disaggregation_hybrid_attention.py
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     popen_launch_pd_server,
 )
 
-register_cuda_ci(est_time=450, suite="stage-c-test-8-gpu-h200")
+register_cuda_ci(est_time=317, suite="stage-c-test-8-gpu-h200")
 
 
 @unittest.skipIf(is_in_ci(), "Temporarily disable the flaky test.")
diff --git a/test/registered/distributed/test_dp_attention.py b/test/registered/distributed/test_dp_attention.py
index 4767d4355..4b9ab0ef5 100644
--- a/test/registered/distributed/test_dp_attention.py
+++ b/test/registered/distributed/test_dp_attention.py
@@ -25,7 +25,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=450, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=446, suite="stage-b-test-2-gpu-large")
 
 
 class TestDPAttentionDP2TP2(
diff --git a/test/registered/distributed/test_epd_disaggregation.py b/test/registered/distributed/test_epd_disaggregation.py
index 301128e19..a4bf5f2c6 100644
--- a/test/registered/distributed/test_epd_disaggregation.py
+++ b/test/registered/distributed/test_epd_disaggregation.py
@@ -36,7 +36,7 @@ DEFAULT_OMNI_MODEL = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
 QWEN35_27B_MODEL = "Qwen/Qwen3.5-27B"
 
 
-register_cuda_ci(est_time=150, suite="stage-c-test-4-gpu-h100")
+register_cuda_ci(est_time=97, suite="stage-c-test-4-gpu-h100")
 
 
 @unittest.skipIf(
diff --git a/test/registered/distributed/test_load_weights_from_remote_instance.py b/test/registered/distributed/test_load_weights_from_remote_instance.py
index 7d51101cd..645ecd82c 100644
--- a/test/registered/distributed/test_load_weights_from_remote_instance.py
+++ b/test/registered/distributed/test_load_weights_from_remote_instance.py
@@ -38,7 +38,7 @@ from sglang.utils import terminate_process
 
 mp.set_start_method("spawn", force=True)
 
-register_cuda_ci(est_time=130, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=110, suite="stage-b-test-2-gpu-large")
 register_amd_ci(est_time=72, suite="stage-b-test-2-gpu-large-amd")
 
 
diff --git a/test/registered/distributed/test_parallel_state.py b/test/registered/distributed/test_parallel_state.py
index 900568139..3a9209b42 100644
--- a/test/registered/distributed/test_parallel_state.py
+++ b/test/registered/distributed/test_parallel_state.py
@@ -43,7 +43,7 @@ import pytest
 
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=5, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-small")
 
 # Import the actual parallel_state module
 parallel_state = pytest.importorskip("sglang.srt.distributed.parallel_state")
diff --git a/test/registered/distributed/test_pp_single_node.py b/test/registered/distributed/test_pp_single_node.py
index cfc26b638..4822ab11d 100644
--- a/test/registered/distributed/test_pp_single_node.py
+++ b/test/registered/distributed/test_pp_single_node.py
@@ -31,7 +31,7 @@ from sglang.test.test_utils import (
     run_bench_one_batch_server,
 )
 
-register_cuda_ci(est_time=460, suite="stage-c-test-4-gpu-h100")
+register_cuda_ci(est_time=462, suite="stage-c-test-4-gpu-h100")
 register_amd_ci(est_time=650, suite="stage-c-test-4-gpu-amd")
 
 
diff --git a/test/registered/dllm/test_llada2_mini.py b/test/registered/dllm/test_llada2_mini.py
index b44a8ab5c..139fc2b7a 100644
--- a/test/registered/dllm/test_llada2_mini.py
+++ b/test/registered/dllm/test_llada2_mini.py
@@ -1,6 +1,6 @@
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=110, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=116, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=330, suite="stage-b-test-1-gpu-small-amd")
 
 import unittest
diff --git a/test/registered/ep/test_deepep_large.py b/test/registered/ep/test_deepep_large.py
index 11c18bab8..aa9d37ec2 100644
--- a/test/registered/ep/test_deepep_large.py
+++ b/test/registered/ep/test_deepep_large.py
@@ -15,7 +15,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=490, suite="stage-c-test-deepep-8-gpu-h200")
+register_cuda_ci(est_time=462, suite="stage-c-test-deepep-8-gpu-h200")
 
 DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
 
diff --git a/test/registered/ep/test_deepep_small.py b/test/registered/ep/test_deepep_small.py
index 01e92d0bd..1f4e8a619 100644
--- a/test/registered/ep/test_deepep_small.py
+++ b/test/registered/ep/test_deepep_small.py
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=430, suite="stage-c-test-deepep-4-gpu-h100")
+register_cuda_ci(est_time=432, suite="stage-c-test-deepep-4-gpu-h100")
 
 
 class TestPureDP(CustomTestCase):
diff --git a/test/registered/ep/test_mooncake_ep_small.py b/test/registered/ep/test_mooncake_ep_small.py
index 37c618991..80f328c0d 100644
--- a/test/registered/ep/test_mooncake_ep_small.py
+++ b/test/registered/ep/test_mooncake_ep_small.py
@@ -15,7 +15,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=80, suite="stage-c-test-deepep-4-gpu-h100")
+register_cuda_ci(est_time=77, suite="stage-c-test-deepep-4-gpu-h100")
 
 ib_devices = get_rdma_devices_args()
 
diff --git a/test/registered/eval/test_eval_accuracy_large.py b/test/registered/eval/test_eval_accuracy_large.py
index 97b441f41..7ac092d8a 100644
--- a/test/registered/eval/test_eval_accuracy_large.py
+++ b/test/registered/eval/test_eval_accuracy_large.py
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=580, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=534, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=420, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/hicache/test_hicache_storage.py b/test/registered/hicache/test_hicache_storage.py
index d6470b198..112178a76 100644
--- a/test/registered/hicache/test_hicache_storage.py
+++ b/test/registered/hicache/test_hicache_storage.py
@@ -1,6 +1,6 @@
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=96, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=99, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=300, suite="stage-b-test-1-gpu-small-amd")
 
 import time
diff --git a/test/registered/hicache/test_hicache_storage_3fs_backend.py b/test/registered/hicache/test_hicache_storage_3fs_backend.py
index 6c478767a..8bf69623d 100644
--- a/test/registered/hicache/test_hicache_storage_3fs_backend.py
+++ b/test/registered/hicache/test_hicache_storage_3fs_backend.py
@@ -13,7 +13,7 @@ from test_hicache_storage_file_backend import HiCacheStorageBaseMixin
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=200, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=150, suite="stage-b-test-2-gpu-large")
 register_amd_ci(est_time=300, suite="stage-b-test-2-gpu-large")
 
 
diff --git a/test/registered/hicache/test_hicache_storage_file_backend.py b/test/registered/hicache/test_hicache_storage_file_backend.py
index 12f779412..99fd26b40 100644
--- a/test/registered/hicache/test_hicache_storage_file_backend.py
+++ b/test/registered/hicache/test_hicache_storage_file_backend.py
@@ -31,7 +31,7 @@ from sglang.test.test_utils import (
 )
 from sglang.utils import wait_for_http_ready
 
-register_cuda_ci(est_time=200, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=148, suite="stage-b-test-2-gpu-large")
 register_amd_ci(est_time=526, suite="stage-b-test-2-gpu-large-amd")
 
 
diff --git a/test/registered/hicache/test_hicache_storage_mooncake_backend.py b/test/registered/hicache/test_hicache_storage_mooncake_backend.py
index 5d928babd..1a9b75a39 100644
--- a/test/registered/hicache/test_hicache_storage_mooncake_backend.py
+++ b/test/registered/hicache/test_hicache_storage_mooncake_backend.py
@@ -20,7 +20,7 @@ from sglang.test.test_utils import (
     is_in_ci,
 )
 
-register_cuda_ci(est_time=230, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=236, suite="stage-b-test-2-gpu-large")
 
 
 class HiCacheStorageMooncakeBackendBaseMixin(HiCacheStorageBaseMixin):
diff --git a/test/registered/hicache/test_hicache_storage_runtime_attach_detach.py b/test/registered/hicache/test_hicache_storage_runtime_attach_detach.py
index 3c9f07224..1517d6d4f 100644
--- a/test/registered/hicache/test_hicache_storage_runtime_attach_detach.py
+++ b/test/registered/hicache/test_hicache_storage_runtime_attach_detach.py
@@ -28,7 +28,7 @@ from sglang.test.test_utils import (
 )
 from sglang.utils import wait_for_http_ready
 
-register_cuda_ci(est_time=130, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=139, suite="stage-b-test-2-gpu-large")
 
 
 class TestHiCacheStorageRuntimeAttachDetach(CustomTestCase):
diff --git a/test/registered/hicache/test_hicache_variants.py b/test/registered/hicache/test_hicache_variants.py
index 956d61d41..c769cf40d 100644
--- a/test/registered/hicache/test_hicache_variants.py
+++ b/test/registered/hicache/test_hicache_variants.py
@@ -1,6 +1,6 @@
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=410, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=450, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=524, suite="stage-b-test-1-gpu-small-amd")
 """
 Consolidated HiCache variant tests.
diff --git a/test/registered/input_embedding/test_input_embeddings.py b/test/registered/input_embedding/test_input_embeddings.py
index cbfc3dc19..373be99dd 100644
--- a/test/registered/input_embedding/test_input_embeddings.py
+++ b/test/registered/input_embedding/test_input_embeddings.py
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=38, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=41, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=38, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/input_embedding/test_input_embeds_chunked.py b/test/registered/input_embedding/test_input_embeds_chunked.py
index 13fc92cea..963a205ce 100644
--- a/test/registered/input_embedding/test_input_embeds_chunked.py
+++ b/test/registered/input_embedding/test_input_embeds_chunked.py
@@ -30,7 +30,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=45, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=42, suite="stage-b-test-1-gpu-small")
 
 CHUNKED_PREFILL_SIZE = 256
 
diff --git a/test/registered/language/test_srt_backend.py b/test/registered/language/test_srt_backend.py
index 24278691c..551ec8103 100644
--- a/test/registered/language/test_srt_backend.py
+++ b/test/registered/language/test_srt_backend.py
@@ -20,7 +20,7 @@ from sglang.test.test_programs import (
 )
 from sglang.test.test_utils import DEFAULT_MODEL_NAME_FOR_TEST, CustomTestCase
 
-register_cuda_ci(est_time=80, suite="stage-a-test-1-gpu-small")
+register_cuda_ci(est_time=72, suite="stage-a-test-1-gpu-small")
 register_amd_ci(est_time=120, suite="stage-a-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/layers/mamba/test_causal_conv1d.py b/test/registered/layers/mamba/test_causal_conv1d.py
index 820a6ff6f..bd10fe21a 100644
--- a/test/registered/layers/mamba/test_causal_conv1d.py
+++ b/test/registered/layers/mamba/test_causal_conv1d.py
@@ -1,6 +1,6 @@
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=25, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=13, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=25, suite="stage-b-test-1-gpu-small-amd")
 
 # Adapted from https://github.com/vllm-project/vllm/blob/main/tests/kernels/mamba/test_causal_conv1d.py
diff --git a/test/registered/layers/mamba/test_mamba2_mixer.py b/test/registered/layers/mamba/test_mamba2_mixer.py
index d63179c1a..85aadbcd7 100644
--- a/test/registered/layers/mamba/test_mamba2_mixer.py
+++ b/test/registered/layers/mamba/test_mamba2_mixer.py
@@ -15,7 +15,7 @@ from sglang.srt.distributed.parallel_state import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=50, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=28, suite="stage-b-test-2-gpu-large")
 
 NUM_GPUS = 2
 
diff --git a/test/registered/layers/mamba/test_mamba_ssm.py b/test/registered/layers/mamba/test_mamba_ssm.py
index 9847684a5..8c139fbdf 100644
--- a/test/registered/layers/mamba/test_mamba_ssm.py
+++ b/test/registered/layers/mamba/test_mamba_ssm.py
@@ -1,6 +1,6 @@
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=7, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=9, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=20, suite="stage-b-test-1-gpu-small-amd")
 
 # Adapted from https://github.com/vllm-project/vllm/blob/633f943e30a4444d890d26b81850f7217736f840/tests/kernels/mamba/test_mamba_ssm_ssd.py
diff --git a/test/registered/layers/mamba/test_mamba_ssm_ssd.py b/test/registered/layers/mamba/test_mamba_ssm_ssd.py
index 670bceabb..2e42caa9b 100644
--- a/test/registered/layers/mamba/test_mamba_ssm_ssd.py
+++ b/test/registered/layers/mamba/test_mamba_ssm_ssd.py
@@ -1,6 +1,6 @@
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=15, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=9, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=34, suite="stage-b-test-1-gpu-small-amd")
 
 # Adapted from https://github.com/vllm-project/vllm/blob/633f943e30a4444d890d26b81850f7217736f840/tests/kernels/mamba/test_mamba_ssm_ssd.py
diff --git a/test/registered/lora/test_fused_moe_lora_kernel.py b/test/registered/lora/test_fused_moe_lora_kernel.py
index 44179a5fa..50587579d 100644
--- a/test/registered/lora/test_fused_moe_lora_kernel.py
+++ b/test/registered/lora/test_fused_moe_lora_kernel.py
@@ -15,7 +15,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
 
 # ==============================================================================
 
-register_cuda_ci(est_time=25, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=14, suite="stage-b-test-1-gpu-large")
 
 
 def round_up(x, base):
diff --git a/test/registered/lora/test_lora_backend.py b/test/registered/lora/test_lora_backend.py
index 5d76cfced..9ac9f4167 100644
--- a/test/registered/lora/test_lora_backend.py
+++ b/test/registered/lora/test_lora_backend.py
@@ -29,7 +29,7 @@ from sglang.test.lora_utils import (
 )
 from sglang.test.test_utils import CustomTestCase, is_in_ci
 
-register_cuda_ci(est_time=200, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=206, suite="stage-b-test-1-gpu-small")
 register_amd_ci(
     est_time=200,
     suite="stage-b-test-1-gpu-small-amd",
diff --git a/test/registered/lora/test_lora_eviction.py b/test/registered/lora/test_lora_eviction.py
index 5a765e9f9..c7c2a20d3 100644
--- a/test/registered/lora/test_lora_eviction.py
+++ b/test/registered/lora/test_lora_eviction.py
@@ -23,7 +23,7 @@ from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.runners import SRTRunner
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=224, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=238, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=224, suite="stage-b-test-1-gpu-small-amd")
 
 PROMPTS = [
diff --git a/test/registered/lora/test_lora_overlap_loading.py b/test/registered/lora/test_lora_overlap_loading.py
index a35b5422b..3812b8c25 100644
--- a/test/registered/lora/test_lora_overlap_loading.py
+++ b/test/registered/lora/test_lora_overlap_loading.py
@@ -29,7 +29,7 @@ from sglang.test.lora_utils import (
 )
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=75, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=45, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=75, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/lora/test_multi_lora_backend.py b/test/registered/lora/test_multi_lora_backend.py
index a0b5db62b..ac0156bbd 100644
--- a/test/registered/lora/test_multi_lora_backend.py
+++ b/test/registered/lora/test_multi_lora_backend.py
@@ -25,7 +25,7 @@ from sglang.test.lora_utils import (
 )
 from sglang.test.test_utils import CustomTestCase, is_in_ci
 
-register_cuda_ci(est_time=100, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=88, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=100, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/mla/test_flashmla.py b/test/registered/mla/test_flashmla.py
index c5f084e42..edd26c2e5 100644
--- a/test/registered/mla/test_flashmla.py
+++ b/test/registered/mla/test_flashmla.py
@@ -21,7 +21,7 @@ from sglang.test.test_utils import (
 )
 
 # FlashMLA attention backend tests with MTP speculative decoding
-register_cuda_ci(est_time=284, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=296, suite="stage-b-test-1-gpu-large")
 
 
 class TestFlashMLAAttnBackend(unittest.TestCase):
diff --git a/test/registered/mla/test_mla.py b/test/registered/mla/test_mla.py
index 03fce55f4..0d6761bdb 100644
--- a/test/registered/mla/test_mla.py
+++ b/test/registered/mla/test_mla.py
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
 )
 
 # MLA attention test with MGSM evaluation
-register_cuda_ci(est_time=194, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=174, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=1100, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/mla/test_mla_deepseek_v3.py b/test/registered/mla/test_mla_deepseek_v3.py
index e1cf2ba83..c0a56c9e3 100644
--- a/test/registered/mla/test_mla_deepseek_v3.py
+++ b/test/registered/mla/test_mla_deepseek_v3.py
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
 )
 
 # DeepSeek-V3 MLA tests with torch compile, FA3, and MTP speculative decoding
-register_cuda_ci(est_time=442, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=543, suite="stage-b-test-1-gpu-large")
 register_amd_ci(
     est_time=221,
     suite="stage-b-test-1-gpu-small-amd",
diff --git a/test/registered/mla/test_mla_flashinfer.py b/test/registered/mla/test_mla_flashinfer.py
index 5bd665805..23a24aade 100644
--- a/test/registered/mla/test_mla_flashinfer.py
+++ b/test/registered/mla/test_mla_flashinfer.py
@@ -15,7 +15,7 @@ from sglang.test.test_utils import (
 )
 
 # FlashInfer MLA backend tests with MTP speculative decoding
-register_cuda_ci(est_time=302, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=256, suite="stage-b-test-1-gpu-large")
 
 
 class TestFlashinferMLA(CustomTestCase):
diff --git a/test/registered/mla/test_mla_fp8.py b/test/registered/mla/test_mla_fp8.py
index a256aad9d..62d27feb6 100644
--- a/test/registered/mla/test_mla_fp8.py
+++ b/test/registered/mla/test_mla_fp8.py
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
 )
 
 # MLA FP8 KV cache test with MGSM evaluation
-register_cuda_ci(est_time=77, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=106, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=800, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/mla/test_mla_int8_deepseek_v3.py b/test/registered/mla/test_mla_int8_deepseek_v3.py
index 1eb00dcce..35cf4b229 100644
--- a/test/registered/mla/test_mla_int8_deepseek_v3.py
+++ b/test/registered/mla/test_mla_int8_deepseek_v3.py
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
 )
 
 # DeepSeek-V3 INT8 quantization tests (channel and block INT8)
-register_cuda_ci(est_time=290, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=282, suite="stage-b-test-1-gpu-large")
 
 
 class TestMLADeepseekV3ChannelInt8(CustomTestCase):
diff --git a/test/registered/model_loading/test_external_models.py b/test/registered/model_loading/test_external_models.py
index c4878abe6..d32adfcc9 100644
--- a/test/registered/model_loading/test_external_models.py
+++ b/test/registered/model_loading/test_external_models.py
@@ -5,7 +5,7 @@ from sglang.srt.environ import envs
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=30, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=28, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=45, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/model_loading/test_utils_update_weights.py b/test/registered/model_loading/test_utils_update_weights.py
index 60f3800b0..f79b6306c 100644
--- a/test/registered/model_loading/test_utils_update_weights.py
+++ b/test/registered/model_loading/test_utils_update_weights.py
@@ -12,7 +12,7 @@ from sglang.srt.weight_sync.utils import update_weights
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import DEFAULT_SMALL_MODEL_NAME_FOR_TEST
 
-register_cuda_ci(est_time=29, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=32, suite="stage-b-test-1-gpu-large")
 
 
 class AsyncEngine(Engine):
diff --git a/test/registered/models/test_compressed_tensors_models.py b/test/registered/models/test_compressed_tensors_models.py
index 3ffc328b2..2d191f40e 100644
--- a/test/registered/models/test_compressed_tensors_models.py
+++ b/test/registered/models/test_compressed_tensors_models.py
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=42, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=63, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=42, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/models/test_generation_models.py b/test/registered/models/test_generation_models.py
index 98e2bac4b..116fc62dc 100644
--- a/test/registered/models/test_generation_models.py
+++ b/test/registered/models/test_generation_models.py
@@ -1,7 +1,7 @@
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
 # Generation model tests (CUDA only)
-register_cuda_ci(est_time=103, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=124, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=106, suite="stage-b-test-1-gpu-small-amd")
 
 # Copyright 2023-2024 SGLang Team
diff --git a/test/registered/models/test_kimi_linear_models.py b/test/registered/models/test_kimi_linear_models.py
index a97a3af8c..83f01d009 100644
--- a/test/registered/models/test_kimi_linear_models.py
+++ b/test/registered/models/test_kimi_linear_models.py
@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=180, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=168, suite="stage-b-test-2-gpu-large")
 
 
 class TestKimiLinear(CustomTestCase):
diff --git a/test/registered/models/test_nvidia_nemotron_3_nano.py b/test/registered/models/test_nvidia_nemotron_3_nano.py
index aba13bf7f..75933f4a4 100644
--- a/test/registered/models/test_nvidia_nemotron_3_nano.py
+++ b/test/registered/models/test_nvidia_nemotron_3_nano.py
@@ -4,7 +4,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.kits.lm_eval_kit import LMEvalMixin
 from sglang.test.server_fixtures.default_fixture import DefaultServerBase
 
-register_cuda_ci(est_time=540, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=564, suite="stage-b-test-2-gpu-large")
 
 NEMOTRON_3_NANO_THINKING_ARGS = [
     "--trust-remote-code",
diff --git a/test/registered/models/test_nvidia_nemotron_nano_v2.py b/test/registered/models/test_nvidia_nemotron_nano_v2.py
index b12395a6a..d9e5883c5 100644
--- a/test/registered/models/test_nvidia_nemotron_nano_v2.py
+++ b/test/registered/models/test_nvidia_nemotron_nano_v2.py
@@ -5,7 +5,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
 from sglang.test.server_fixtures.default_fixture import DefaultServerBase
 
-register_cuda_ci(est_time=240, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=214, suite="stage-b-test-2-gpu-large")
 
 
 class TestNvidiaNemotronNanoV2BF16(GSM8KMixin, DefaultServerBase):
diff --git a/test/registered/models/test_nvidia_nemotron_nano_v2_vl.py b/test/registered/models/test_nvidia_nemotron_nano_v2_vl.py
index 71211b7ec..4c52c05d6 100644
--- a/test/registered/models/test_nvidia_nemotron_nano_v2_vl.py
+++ b/test/registered/models/test_nvidia_nemotron_nano_v2_vl.py
@@ -10,7 +10,7 @@ from sglang.test.server_fixtures.mmmu_fixture import MMMUServerBase
 # GSM8k + MMMU evaluation
 
 
-register_cuda_ci(est_time=214, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=206, suite="stage-b-test-1-gpu-large")
 
 MODEL = "nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16"
 
diff --git a/test/registered/models/test_qwen_models.py b/test/registered/models/test_qwen_models.py
index 24817c816..602de7f25 100644
--- a/test/registered/models/test_qwen_models.py
+++ b/test/registered/models/test_qwen_models.py
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=90, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=109, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=130, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/models/test_transformers_backend_eval.py b/test/registered/models/test_transformers_backend_eval.py
index 16125aa57..9d68505fa 100644
--- a/test/registered/models/test_transformers_backend_eval.py
+++ b/test/registered/models/test_transformers_backend_eval.py
@@ -7,7 +7,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.few_shot_gsm8k import run_eval
 from sglang.test.server_fixtures.default_fixture import DefaultServerBase
 
-register_cuda_ci(est_time=40, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=44, suite="stage-b-test-1-gpu-small")
 
 
 class TestTransformersBackendEval(DefaultServerBase):
diff --git a/test/registered/models/test_transformers_models.py b/test/registered/models/test_transformers_models.py
index 39339f6de..87fc3fbe5 100644
--- a/test/registered/models/test_transformers_models.py
+++ b/test/registered/models/test_transformers_models.py
@@ -21,7 +21,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=260, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=244, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=320, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/models/test_vlm_models.py b/test/registered/models/test_vlm_models.py
index 7789bed9d..2ed573417 100644
--- a/test/registered/models/test_vlm_models.py
+++ b/test/registered/models/test_vlm_models.py
@@ -13,7 +13,7 @@ from sglang.test.test_utils import is_in_ci
 # VLM (Vision Language Model) tests
 
 
-register_cuda_ci(est_time=228, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=176, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=850, suite="stage-b-test-1-gpu-small-amd-nondeterministic")
 
 _is_hip = is_hip()
diff --git a/test/registered/moe/test_cutedsl_moe.py b/test/registered/moe/test_cutedsl_moe.py
index f8449c4c9..baed48c5b 100644
--- a/test/registered/moe/test_cutedsl_moe.py
+++ b/test/registered/moe/test_cutedsl_moe.py
@@ -16,7 +16,7 @@ except ImportError:
     CuteDslMoEWrapper = None
     convert_sf_to_mma_layout = None
 
-register_cuda_ci(est_time=300, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=13, suite="stage-c-test-4-gpu-b200")
 
 SKIP_TEST = torch.cuda.get_device_capability() < (10, 0)
 SKIP_REASON = "Nvfp4 Requires compute capability of 10 or above."
diff --git a/test/registered/moe/test_fused_moe.py b/test/registered/moe/test_fused_moe.py
index d67f90356..2b267fa9e 100644
--- a/test/registered/moe/test_fused_moe.py
+++ b/test/registered/moe/test_fused_moe.py
@@ -13,7 +13,7 @@ from sglang.srt.utils import get_device, get_device_capability, is_hip
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase, empty_gpu_cache
 
-register_cuda_ci(est_time=80, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=79, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=30, suite="stage-b-test-1-gpu-small-amd")
 
 _is_hip = is_hip()
diff --git a/test/registered/moe/test_glm4_moe_models.py b/test/registered/moe/test_glm4_moe_models.py
index 010556b83..58a54f25b 100644
--- a/test/registered/moe/test_glm4_moe_models.py
+++ b/test/registered/moe/test_glm4_moe_models.py
@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=100, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=150, suite="stage-b-test-2-gpu-large")
 
 
 class TestGLM4MoE(CustomTestCase):
diff --git a/test/registered/moe/test_moe_ep.py b/test/registered/moe/test_moe_ep.py
index 1d27d3c20..5aa4d4836 100644
--- a/test/registered/moe/test_moe_ep.py
+++ b/test/registered/moe/test_moe_ep.py
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=250, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=232, suite="stage-b-test-2-gpu-large")
 
 
 class TestEp(CustomTestCase):
diff --git a/test/registered/moe/test_torch_compile_moe.py b/test/registered/moe/test_torch_compile_moe.py
index 998ddc61c..0da421835 100644
--- a/test/registered/moe/test_torch_compile_moe.py
+++ b/test/registered/moe/test_torch_compile_moe.py
@@ -15,7 +15,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=120, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=125, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=1400, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/moe/test_triton_fused_moe.py b/test/registered/moe/test_triton_fused_moe.py
index 2255d64bf..b5c263a04 100644
--- a/test/registered/moe/test_triton_fused_moe.py
+++ b/test/registered/moe/test_triton_fused_moe.py
@@ -12,7 +12,7 @@ from sglang.srt.server_args import ServerArgs, set_global_server_args_for_schedu
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=30, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=12, suite="stage-b-test-1-gpu-large")
 
 
 class TestFusedMOE(CustomTestCase):
diff --git a/test/registered/observability/test_metrics.py b/test/registered/observability/test_metrics.py
index d1f0f381b..066999d02 100644
--- a/test/registered/observability/test_metrics.py
+++ b/test/registered/observability/test_metrics.py
@@ -20,7 +20,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=95, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=73, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=32, suite="stage-b-test-1-gpu-small-amd")
 
 _MODEL_NAME = "Qwen/Qwen3-0.6B"
diff --git a/test/registered/observability/test_tracing.py b/test/registered/observability/test_tracing.py
index 995c3196a..c79b325db 100644
--- a/test/registered/observability/test_tracing.py
+++ b/test/registered/observability/test_tracing.py
@@ -49,7 +49,7 @@ from sglang.test.test_utils import (
 logger = logging.getLogger(__name__)
 
 # CI registration
-register_cuda_ci(est_time=120, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=104, suite="stage-b-test-1-gpu-small")
 
 
 # ============================================================================
diff --git a/test/registered/observability/test_tracing_disaggregation.py b/test/registered/observability/test_tracing_disaggregation.py
index df15ae101..f06d2718d 100644
--- a/test/registered/observability/test_tracing_disaggregation.py
+++ b/test/registered/observability/test_tracing_disaggregation.py
@@ -35,7 +35,7 @@ from sglang.utils import wait_for_http_ready
 logger = logging.getLogger(__name__)
 
 # CI registration - PD disaggregation requires 2 GPUs
-register_cuda_ci(est_time=45, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=48, suite="stage-b-test-2-gpu-large")
 
 
 class TestTraceDisaggregation(CustomTestCase):
diff --git a/test/registered/openai_server/basic/test_anthropic_server.py b/test/registered/openai_server/basic/test_anthropic_server.py
index bdfb8b91f..3e50e6758 100644
--- a/test/registered/openai_server/basic/test_anthropic_server.py
+++ b/test/registered/openai_server/basic/test_anthropic_server.py
@@ -31,7 +31,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=40, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=39, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=140, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/openai_server/basic/test_http2_server.py b/test/registered/openai_server/basic/test_http2_server.py
index 31c21304c..d35204c31 100644
--- a/test/registered/openai_server/basic/test_http2_server.py
+++ b/test/registered/openai_server/basic/test_http2_server.py
@@ -27,7 +27,7 @@ try:
 except ImportError:
     _HAS_GRANIAN = False
 
-register_cuda_ci(est_time=50, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=51, suite="stage-b-test-1-gpu-small")
 
 
 @unittest.skipUnless(_HAS_GRANIAN, "granian not installed (pip install sglang[http2])")
diff --git a/test/registered/openai_server/basic/test_openai_server.py b/test/registered/openai_server/basic/test_openai_server.py
index ce1b8ec51..967ef109b 100644
--- a/test/registered/openai_server/basic/test_openai_server.py
+++ b/test/registered/openai_server/basic/test_openai_server.py
@@ -28,7 +28,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=184, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=189, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=200, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/openai_server/basic/test_protocol.py b/test/registered/openai_server/basic/test_protocol.py
index bdbdff6a1..b707f35bf 100644
--- a/test/registered/openai_server/basic/test_protocol.py
+++ b/test/registered/openai_server/basic/test_protocol.py
@@ -30,7 +30,7 @@ from sglang.srt.entrypoints.openai.protocol import (
 )
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=3, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=2, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=10, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/openai_server/basic/test_serving_chat.py b/test/registered/openai_server/basic/test_serving_chat.py
index 2ca6135ac..e1d765aa8 100644
--- a/test/registered/openai_server/basic/test_serving_chat.py
+++ b/test/registered/openai_server/basic/test_serving_chat.py
@@ -24,7 +24,7 @@ from sglang.srt.managers.io_struct import GenerateReqInput
 from sglang.srt.utils import get_or_create_event_loop
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=10, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=10, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/openai_server/basic/test_serving_completions.py b/test/registered/openai_server/basic/test_serving_completions.py
index 026f1c2d8..1ded89c39 100644
--- a/test/registered/openai_server/basic/test_serving_completions.py
+++ b/test/registered/openai_server/basic/test_serving_completions.py
@@ -18,7 +18,7 @@ from sglang.srt.managers.tokenizer_manager import TokenizerManager
 from sglang.srt.utils import get_or_create_event_loop
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=10, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=10, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/openai_server/features/test_json_mode.py b/test/registered/openai_server/features/test_json_mode.py
index 8cb922ecb..fcdd6e792 100644
--- a/test/registered/openai_server/features/test_json_mode.py
+++ b/test/registered/openai_server/features/test_json_mode.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=109, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=108, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=180, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/openai_server/features/test_openai_server_ebnf.py b/test/registered/openai_server/features/test_openai_server_ebnf.py
index 2ec7d2dcf..78327bc50 100644
--- a/test/registered/openai_server/features/test_openai_server_ebnf.py
+++ b/test/registered/openai_server/features/test_openai_server_ebnf.py
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=55, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=42, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=20, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/openai_server/features/test_openai_server_hidden_states.py b/test/registered/openai_server/features/test_openai_server_hidden_states.py
index ff1070385..868e9aa7a 100644
--- a/test/registered/openai_server/features/test_openai_server_hidden_states.py
+++ b/test/registered/openai_server/features/test_openai_server_hidden_states.py
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=186, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=193, suite="stage-b-test-1-gpu-small")
 register_amd_ci(
     est_time=186,
     suite="stage-b-test-1-gpu-small-amd",
diff --git a/test/registered/openai_server/function_call/test_openai_function_calling.py b/test/registered/openai_server/function_call/test_openai_function_calling.py
index 33cde4018..ea02b8951 100644
--- a/test/registered/openai_server/function_call/test_openai_function_calling.py
+++ b/test/registered/openai_server/function_call/test_openai_function_calling.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=60, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=97, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=73, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/openai_server/function_call/test_tool_choice.py b/test/registered/openai_server/function_call/test_tool_choice.py
index 187cb8827..a20cd4e56 100644
--- a/test/registered/openai_server/function_call/test_tool_choice.py
+++ b/test/registered/openai_server/function_call/test_tool_choice.py
@@ -22,7 +22,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=250, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=194, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=258, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/openai_server/validation/test_large_max_new_tokens.py b/test/registered/openai_server/validation/test_large_max_new_tokens.py
index 981e3af72..c48ca4547 100644
--- a/test/registered/openai_server/validation/test_large_max_new_tokens.py
+++ b/test/registered/openai_server/validation/test_large_max_new_tokens.py
@@ -22,7 +22,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=41, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=56, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=41, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/openai_server/validation/test_matched_stop.py b/test/registered/openai_server/validation/test_matched_stop.py
index 0611a1ea2..da8ec59cf 100644
--- a/test/registered/openai_server/validation/test_matched_stop.py
+++ b/test/registered/openai_server/validation/test_matched_stop.py
@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=40, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=52, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=60, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/openai_server/validation/test_openai_server_ignore_eos.py b/test/registered/openai_server/validation/test_openai_server_ignore_eos.py
index 2f27699ee..2986c2f57 100644
--- a/test/registered/openai_server/validation/test_openai_server_ignore_eos.py
+++ b/test/registered/openai_server/validation/test_openai_server_ignore_eos.py
@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=55, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=42, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=47, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/openai_server/validation/test_request_length_validation.py b/test/registered/openai_server/validation/test_request_length_validation.py
index 0699d3917..1bc93e551 100644
--- a/test/registered/openai_server/validation/test_request_length_validation.py
+++ b/test/registered/openai_server/validation/test_request_length_validation.py
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=38, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=39, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=31, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/ops/test_repeat_interleave.py b/test/registered/ops/test_repeat_interleave.py
index acb45bd47..ca2c26b91 100644
--- a/test/registered/ops/test_repeat_interleave.py
+++ b/test/registered/ops/test_repeat_interleave.py
@@ -11,7 +11,7 @@ from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 # Ops - Repeat Interleave tests (1-GPU)
 
 
-register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=7, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=75, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/perf/test_bench_one_batch_1gpu.py b/test/registered/perf/test_bench_one_batch_1gpu.py
index acc8ccafb..26c251ccd 100644
--- a/test/registered/perf/test_bench_one_batch_1gpu.py
+++ b/test/registered/perf/test_bench_one_batch_1gpu.py
@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=120, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=80, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=120, suite="stage-b-test-1-gpu-large-amd")
 
 
diff --git a/test/registered/perf/test_bench_one_batch_2gpu.py b/test/registered/perf/test_bench_one_batch_2gpu.py
index 38fe4b3c8..41ba7a030 100644
--- a/test/registered/perf/test_bench_one_batch_2gpu.py
+++ b/test/registered/perf/test_bench_one_batch_2gpu.py
@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=180, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=167, suite="stage-b-test-2-gpu-large")
 register_amd_ci(est_time=630, suite="stage-b-test-2-gpu-large-amd")
 
 
diff --git a/test/registered/perf/test_bench_serving_1gpu_large.py b/test/registered/perf/test_bench_serving_1gpu_large.py
index 36f17107d..43e0ad63e 100644
--- a/test/registered/perf/test_bench_serving_1gpu_large.py
+++ b/test/registered/perf/test_bench_serving_1gpu_large.py
@@ -17,7 +17,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=300, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=251, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=300, suite="stage-b-test-1-gpu-large-amd")
 
 
diff --git a/test/registered/perf/test_bench_serving_1gpu_part1.py b/test/registered/perf/test_bench_serving_1gpu_part1.py
index 9e2778c57..fffbd5525 100644
--- a/test/registered/perf/test_bench_serving_1gpu_part1.py
+++ b/test/registered/perf/test_bench_serving_1gpu_part1.py
@@ -19,7 +19,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=1140, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=1184, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=1100, suite="stage-b-test-1-gpu-large-amd")
 
 
diff --git a/test/registered/perf/test_bench_serving_1gpu_part2.py b/test/registered/perf/test_bench_serving_1gpu_part2.py
index 5bb97ef49..9f3725302 100644
--- a/test/registered/perf/test_bench_serving_1gpu_part2.py
+++ b/test/registered/perf/test_bench_serving_1gpu_part2.py
@@ -19,7 +19,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=900, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=910, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=900, suite="stage-b-test-1-gpu-large-amd")
 
 
diff --git a/test/registered/perf/test_bench_serving_2gpu.py b/test/registered/perf/test_bench_serving_2gpu.py
index 78320decf..7da00c13f 100644
--- a/test/registered/perf/test_bench_serving_2gpu.py
+++ b/test/registered/perf/test_bench_serving_2gpu.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=600, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=660, suite="stage-b-test-2-gpu-large")
 register_amd_ci(est_time=1100, suite="stage-b-test-2-gpu-large-amd")
 
 
diff --git a/test/registered/piecewise_cuda_graph/test_piecewise_cuda_graph_support_1_gpu.py b/test/registered/piecewise_cuda_graph/test_piecewise_cuda_graph_support_1_gpu.py
index ce6fe2291..25a542cfd 100644
--- a/test/registered/piecewise_cuda_graph/test_piecewise_cuda_graph_support_1_gpu.py
+++ b/test/registered/piecewise_cuda_graph/test_piecewise_cuda_graph_support_1_gpu.py
@@ -17,7 +17,7 @@ from sglang.test.test_utils import (
 )
 
 # CI Registration
-register_cuda_ci(est_time=220, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=262, suite="stage-b-test-1-gpu-large")
 
 
 class TestPiecewiseCudaGraphQwen25VL(CustomTestCase):
diff --git a/test/registered/prefill_only/test_cross_encoder_models.py b/test/registered/prefill_only/test_cross_encoder_models.py
index 453d67a60..d63323d54 100644
--- a/test/registered/prefill_only/test_cross_encoder_models.py
+++ b/test/registered/prefill_only/test_cross_encoder_models.py
@@ -11,7 +11,7 @@ from sglang.test.test_utils import CustomTestCase, is_in_ci
 # Cross encoder model tests
 
 
-register_cuda_ci(est_time=100, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=101, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=150, suite="stage-b-test-1-gpu-small-amd")
 
 MODELS = [
diff --git a/test/registered/prefill_only/test_embed_overrides.py b/test/registered/prefill_only/test_embed_overrides.py
index 42a77c6a9..7075472e6 100644
--- a/test/registered/prefill_only/test_embed_overrides.py
+++ b/test/registered/prefill_only/test_embed_overrides.py
@@ -23,7 +23,7 @@ from sglang.srt.managers.tokenizer_manager_score_mixin import (
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=5, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-small")
 
 HIDDEN_DIM = 4
 
diff --git a/test/registered/prefill_only/test_embedding_models.py b/test/registered/prefill_only/test_embedding_models.py
index bc8f48059..6776c8602 100644
--- a/test/registered/prefill_only/test_embedding_models.py
+++ b/test/registered/prefill_only/test_embedding_models.py
@@ -35,7 +35,7 @@ register_amd_ci(
     suite="stage-b-test-1-gpu-small-amd",
     disabled="see https://github.com/sgl-project/sglang/issues/11127",
 )
-register_cuda_ci(est_time=90, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=102, suite="stage-b-test-1-gpu-small")
 
 MODEL_TO_CONFIG = {
     "Alibaba-NLP/gte-Qwen2-1.5B-instruct": (1, 1e-5),
diff --git a/test/registered/prefill_only/test_encoder_embedding_models.py b/test/registered/prefill_only/test_encoder_embedding_models.py
index 64d9c5dba..90cfca661 100644
--- a/test/registered/prefill_only/test_encoder_embedding_models.py
+++ b/test/registered/prefill_only/test_encoder_embedding_models.py
@@ -29,7 +29,7 @@ from sglang.test.test_utils import CustomTestCase, get_similarities, is_in_ci
 # python -m unittest test_encoder_embedding_models.TestEncoderEmbeddingModels.test_prefill_logits
 
 
-register_cuda_ci(est_time=310, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=198, suite="stage-b-test-1-gpu-small")
 
 MODELS = [("BAAI/bge-small-en", 1, 1e-5), ("BAAI/bge-m3", 1, 1e-5)]
 
diff --git a/test/registered/prefill_only/test_openai_embedding.py b/test/registered/prefill_only/test_openai_embedding.py
index 3c13198e8..cb3507027 100644
--- a/test/registered/prefill_only/test_openai_embedding.py
+++ b/test/registered/prefill_only/test_openai_embedding.py
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=70, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=80, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=141, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/prefill_only/test_reward_models.py b/test/registered/prefill_only/test_reward_models.py
index a26314e62..fb3084675 100644
--- a/test/registered/prefill_only/test_reward_models.py
+++ b/test/registered/prefill_only/test_reward_models.py
@@ -24,7 +24,7 @@ from sglang.test.test_utils import CustomTestCase
 # ==============================================================================
 
 
-register_cuda_ci(est_time=103, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=142, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=132, suite="stage-b-test-1-gpu-small-amd-nondeterministic")
 
 MODELS = [
diff --git a/test/registered/prefill_only/test_score_api.py b/test/registered/prefill_only/test_score_api.py
index b1072ef34..c096290d6 100644
--- a/test/registered/prefill_only/test_score_api.py
+++ b/test/registered/prefill_only/test_score_api.py
@@ -26,7 +26,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=160, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=70, suite="stage-b-test-1-gpu-small")
 
 _MODEL = DEFAULT_SMALL_MODEL_NAME_FOR_TEST  # Llama-3.2-1B-Instruct
 # <|eot_id|> for Llama-3.x Instruct — used as MIS delimiter
diff --git a/test/registered/prefill_only/test_score_engine.py b/test/registered/prefill_only/test_score_engine.py
index a282c1a05..767e8a91e 100644
--- a/test/registered/prefill_only/test_score_engine.py
+++ b/test/registered/prefill_only/test_score_engine.py
@@ -22,7 +22,7 @@ from sglang.srt.entrypoints.engine import Engine
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import DEFAULT_SMALL_MODEL_NAME_FOR_TEST, CustomTestCase
 
-register_cuda_ci(est_time=200, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=85, suite="stage-b-test-1-gpu-small")
 
 _CAUSAL_LM_MODEL = DEFAULT_SMALL_MODEL_NAME_FOR_TEST  # Llama-3.2-1B-Instruct
 _SEQCLS_MODEL = "Qwen/Qwen3-0.6B"  # backbone; arch overridden to SeqCls below
diff --git a/test/registered/prefill_only/test_serving_rerank.py b/test/registered/prefill_only/test_serving_rerank.py
index edfb55657..411ea1c8e 100644
--- a/test/registered/prefill_only/test_serving_rerank.py
+++ b/test/registered/prefill_only/test_serving_rerank.py
@@ -7,7 +7,7 @@ from sglang.srt.managers.tokenizer_manager_score_mixin import ScoreResult
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
 # Keep consistent with other openai_server/basic unit tests.
-register_cuda_ci(est_time=10, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=10, suite="stage-b-test-1-gpu-small-amd")
 
 try:
diff --git a/test/registered/profiling/test_start_profile.py b/test/registered/profiling/test_start_profile.py
index a2d25e8e9..edfb48fa5 100644
--- a/test/registered/profiling/test_start_profile.py
+++ b/test/registered/profiling/test_start_profile.py
@@ -29,7 +29,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=41, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=40, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=60, suite="stage-b-test-1-gpu-small-amd")
 
 OUTPUT_DIR = "./profiler_dir"
diff --git a/test/registered/quant/test_autoround.py b/test/registered/quant/test_autoround.py
index 89374ec56..9c1e3005f 100644
--- a/test/registered/quant/test_autoround.py
+++ b/test/registered/quant/test_autoround.py
@@ -17,7 +17,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=77, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=96, suite="stage-b-test-1-gpu-large")
 
 
 class TestAutoRound(CustomTestCase):
diff --git a/test/registered/quant/test_awq.py b/test/registered/quant/test_awq.py
index c2a267652..b94fe4006 100644
--- a/test/registered/quant/test_awq.py
+++ b/test/registered/quant/test_awq.py
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=260, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=209, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=200, suite="stage-b-test-1-gpu-large-amd")
 
 
diff --git a/test/registered/quant/test_block_int8.py b/test/registered/quant/test_block_int8.py
index 68ddf2fc9..68829f90f 100644
--- a/test/registered/quant/test_block_int8.py
+++ b/test/registered/quant/test_block_int8.py
@@ -10,7 +10,7 @@ from sglang.srt.server_args import ServerArgs, set_global_server_args_for_schedu
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=44, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=39, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=22, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/quant/test_bnb.py b/test/registered/quant/test_bnb.py
index 8ce583f22..ee7964256 100644
--- a/test/registered/quant/test_bnb.py
+++ b/test/registered/quant/test_bnb.py
@@ -22,7 +22,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=5, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=6, suite="stage-b-test-1-gpu-small")
 
 VISION_MODELS = [
     "unsloth/Qwen2.5-VL-7B-Instruct-bnb-4bit",
diff --git a/test/registered/quant/test_deepseek_v32_fp4_4gpu.py b/test/registered/quant/test_deepseek_v32_fp4_4gpu.py
index f618d8332..dc5d2824b 100644
--- a/test/registered/quant/test_deepseek_v32_fp4_4gpu.py
+++ b/test/registered/quant/test_deepseek_v32_fp4_4gpu.py
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=530, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=852, suite="stage-c-test-4-gpu-b200")
 
 FULL_DEEPSEEK_V3_FP4_MODEL_PATH = "nvidia/DeepSeek-V3.2-NVFP4"
 SERVER_LAUNCH_TIMEOUT = 1200
diff --git a/test/registered/quant/test_deepseek_v32_fp4_mtp_4gpu.py b/test/registered/quant/test_deepseek_v32_fp4_mtp_4gpu.py
index db2298423..8ea65b8ca 100644
--- a/test/registered/quant/test_deepseek_v32_fp4_mtp_4gpu.py
+++ b/test/registered/quant/test_deepseek_v32_fp4_mtp_4gpu.py
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=610, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=1030, suite="stage-c-test-4-gpu-b200")
 
 FULL_DEEPSEEK_V3_FP4_MODEL_PATH = "nvidia/DeepSeek-V3.2-NVFP4"
 SERVER_LAUNCH_TIMEOUT = 1200
diff --git a/test/registered/quant/test_deepseek_v3_fp4_4gpu.py b/test/registered/quant/test_deepseek_v3_fp4_4gpu.py
index 3a9030cfd..ed32e53aa 100644
--- a/test/registered/quant/test_deepseek_v3_fp4_4gpu.py
+++ b/test/registered/quant/test_deepseek_v3_fp4_4gpu.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=450, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=1146, suite="stage-c-test-4-gpu-b200")
 
 FULL_DEEPSEEK_V3_FP4_MODEL_PATH = "nvidia/DeepSeek-V3-0324-FP4"
 SERVER_LAUNCH_TIMEOUT = 1200
diff --git a/test/registered/quant/test_eval_fp8_accuracy.py b/test/registered/quant/test_eval_fp8_accuracy.py
index f60fcc7af..5fa425a24 100644
--- a/test/registered/quant/test_eval_fp8_accuracy.py
+++ b/test/registered/quant/test_eval_fp8_accuracy.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=250, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=326, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=600, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/quant/test_fp8_blockwise_gemm.py b/test/registered/quant/test_fp8_blockwise_gemm.py
index a9602d94c..aace46e55 100644
--- a/test/registered/quant/test_fp8_blockwise_gemm.py
+++ b/test/registered/quant/test_fp8_blockwise_gemm.py
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=400, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=302, suite="stage-c-test-4-gpu-b200")
 
 MODEL_PATH = "Qwen/Qwen3-4B-Instruct-2507-FP8"
 MXFP8_MODEL_PATH = "zianglih/Qwen3-4B-Instruct-2507-MXFP8"
diff --git a/test/registered/quant/test_fp8_gemm_sm120.py b/test/registered/quant/test_fp8_gemm_sm120.py
index 9ca9ea7af..bbdf97fa6 100644
--- a/test/registered/quant/test_fp8_gemm_sm120.py
+++ b/test/registered/quant/test_fp8_gemm_sm120.py
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=120, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=144, suite="stage-b-test-1-gpu-small")
 
 PERTENSOR_MODEL_PATH = "nvidia/Llama-3.1-8B-Instruct-FP8"
 BLOCKWISE_MODEL_PATH = "Qwen/Qwen3-4B-Instruct-2507-FP8"
diff --git a/test/registered/quant/test_fp8_kernel.py b/test/registered/quant/test_fp8_kernel.py
index ea5c60689..5ae5c0485 100644
--- a/test/registered/quant/test_fp8_kernel.py
+++ b/test/registered/quant/test_fp8_kernel.py
@@ -9,7 +9,7 @@ from sglang.srt.layers.quantization.fp8_kernel import (
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=15, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=10, suite="stage-b-test-1-gpu-large")
 
 
 class TestFP8Base(CustomTestCase):
diff --git a/test/registered/quant/test_fp8_utils.py b/test/registered/quant/test_fp8_utils.py
index 563a9123b..b4118d6f6 100644
--- a/test/registered/quant/test_fp8_utils.py
+++ b/test/registered/quant/test_fp8_utils.py
@@ -10,7 +10,7 @@ from sglang.srt.layers.quantization.fp8_utils import (
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=9, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-large")
 
 
 class TestInverseTransformScaleUe8m0(CustomTestCase):
diff --git a/test/registered/quant/test_fp8kv_triton.py b/test/registered/quant/test_fp8kv_triton.py
index df9573e22..8136f0109 100644
--- a/test/registered/quant/test_fp8kv_triton.py
+++ b/test/registered/quant/test_fp8kv_triton.py
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=80, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=63, suite="stage-b-test-1-gpu-large")
 
 
 class TestFP8KVCacheTritonBackend(CustomTestCase):
diff --git a/test/registered/quant/test_gguf.py b/test/registered/quant/test_gguf.py
index 853d3dda7..4ae3bb8e0 100644
--- a/test/registered/quant/test_gguf.py
+++ b/test/registered/quant/test_gguf.py
@@ -6,7 +6,7 @@ import sglang as sgl
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=96, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=71, suite="stage-b-test-1-gpu-small")
 
 
 class TestGGUF(CustomTestCase):
diff --git a/test/registered/quant/test_gptqmodel_dynamic.py b/test/registered/quant/test_gptqmodel_dynamic.py
index 6fdbf676f..15927a7c9 100644
--- a/test/registered/quant/test_gptqmodel_dynamic.py
+++ b/test/registered/quant/test_gptqmodel_dynamic.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=102, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=98, suite="stage-b-test-1-gpu-large")
 
 
 def check_quant_method(model_path: str, use_marlin_kernel: bool):
diff --git a/test/registered/quant/test_int8_kernel.py b/test/registered/quant/test_int8_kernel.py
index 0a5d3001a..84591e3b4 100644
--- a/test/registered/quant/test_int8_kernel.py
+++ b/test/registered/quant/test_int8_kernel.py
@@ -11,7 +11,7 @@ from sglang.srt.server_args import ServerArgs, set_global_server_args_for_schedu
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=16, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=13, suite="stage-b-test-1-gpu-small")
 
 
 def native_w8a8_per_token_matmul(A, B, As, Bs, output_dtype=torch.float16):
diff --git a/test/registered/quant/test_marlin_moe.py b/test/registered/quant/test_marlin_moe.py
index 9a043882f..9554c51f6 100644
--- a/test/registered/quant/test_marlin_moe.py
+++ b/test/registered/quant/test_marlin_moe.py
@@ -12,7 +12,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_marlin_utils import awq_marlin_quantize, marlin_quantize
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=130, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=85, suite="stage-b-test-1-gpu-small")
 
 set_global_server_args_for_scheduler(object.__new__(ServerArgs))
 
diff --git a/test/registered/quant/test_modelopt_fp8.py b/test/registered/quant/test_modelopt_fp8.py
index a65e2e20c..9f1c7bdfc 100644
--- a/test/registered/quant/test_modelopt_fp8.py
+++ b/test/registered/quant/test_modelopt_fp8.py
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=120, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=62, suite="stage-b-test-1-gpu-large")
 
 
 class TestModeloptFP8(CustomTestCase):
diff --git a/test/registered/quant/test_nvfp4_gemm.py b/test/registered/quant/test_nvfp4_gemm.py
index dacbb04b9..7325231e8 100644
--- a/test/registered/quant/test_nvfp4_gemm.py
+++ b/test/registered/quant/test_nvfp4_gemm.py
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=300, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=322, suite="stage-c-test-4-gpu-b200")
 
 MODEL_PATH = "nvidia/Llama-3.1-8B-Instruct-NVFP4"
 
diff --git a/test/registered/quant/test_nvfp4_gemm_sm120.py b/test/registered/quant/test_nvfp4_gemm_sm120.py
index dea1e17cc..4f87b965b 100644
--- a/test/registered/quant/test_nvfp4_gemm_sm120.py
+++ b/test/registered/quant/test_nvfp4_gemm_sm120.py
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=90, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=115, suite="stage-b-test-1-gpu-small")
 
 MODEL_PATH = "nvidia/Llama-3.1-8B-Instruct-NVFP4"
 
diff --git a/test/registered/quant/test_quant_config_parsing.py b/test/registered/quant/test_quant_config_parsing.py
index 5b9d12817..5db3be7b8 100644
--- a/test/registered/quant/test_quant_config_parsing.py
+++ b/test/registered/quant/test_quant_config_parsing.py
@@ -5,7 +5,7 @@ from sglang.srt.configs.model_config import ModelConfig
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cpu_ci(est_time=20, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=12, suite="stage-a-test-cpu")
 
 
 class TestQuantLogString(CustomTestCase):
diff --git a/test/registered/quant/test_quantization.py b/test/registered/quant/test_quantization.py
index 083282b87..30ccc18f8 100644
--- a/test/registered/quant/test_quantization.py
+++ b/test/registered/quant/test_quantization.py
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
     write_results_to_json,
 )
 
-register_cuda_ci(est_time=270, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=402, suite="stage-b-test-1-gpu-large")
 
 MODEL_SCORE_THRESHOLDS = {
     # Baselines observed with gsm8k 5-shot concatenated format via chat API,
diff --git a/test/registered/quant/test_torchao.py b/test/registered/quant/test_torchao.py
index 16dc187a7..018f1a02e 100644
--- a/test/registered/quant/test_torchao.py
+++ b/test/registered/quant/test_torchao.py
@@ -5,7 +5,7 @@ import requests
 from sglang import Engine
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=140, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=124, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=230, suite="stage-b-test-1-gpu-small-amd")
 from sglang.lang.chat_template import get_chat_template_by_model_path
 from sglang.srt.utils import kill_process_tree
diff --git a/test/registered/quant/test_triton_scaled_mm.py b/test/registered/quant/test_triton_scaled_mm.py
index 2a4010240..f66c33c5d 100644
--- a/test/registered/quant/test_triton_scaled_mm.py
+++ b/test/registered/quant/test_triton_scaled_mm.py
@@ -8,7 +8,7 @@ from sglang.srt.layers.quantization.fp8_kernel import triton_scaled_mm
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=9, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=12, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/quant/test_w8a8_quantization.py b/test/registered/quant/test_w8a8_quantization.py
index a2a2a1cb4..33805704e 100644
--- a/test/registered/quant/test_w8a8_quantization.py
+++ b/test/registered/quant/test_w8a8_quantization.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=160, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=204, suite="stage-b-test-1-gpu-large")
 
 
 class BaseW8A8Test(CustomTestCase):
diff --git a/test/registered/radix_cache/test_radix_attention.py b/test/registered/radix_cache/test_radix_attention.py
index 871dac26e..3ce4643da 100644
--- a/test/registered/radix_cache/test_radix_attention.py
+++ b/test/registered/radix_cache/test_radix_attention.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
 )
 
 # RadixAttention server integration tests
-register_cuda_ci(est_time=100, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=96, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=100, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/radix_cache/test_radix_cache_hit.py b/test/registered/radix_cache/test_radix_cache_hit.py
index 43f63cad7..974adb866 100644
--- a/test/registered/radix_cache/test_radix_cache_hit.py
+++ b/test/registered/radix_cache/test_radix_cache_hit.py
@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=120, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=62, suite="stage-b-test-1-gpu-small")
 
 MODEL = DEFAULT_SMALL_MODEL_NAME_FOR_TEST
 
diff --git a/test/registered/radix_cache/test_swa_radix_cache_kl.py b/test/registered/radix_cache/test_swa_radix_cache_kl.py
index db60d0055..a58abd916 100644
--- a/test/registered/radix_cache/test_swa_radix_cache_kl.py
+++ b/test/registered/radix_cache/test_swa_radix_cache_kl.py
@@ -6,7 +6,7 @@ from sglang.test.server_fixtures.default_fixture import DefaultServerBase
 
 MODEL = "openai/gpt-oss-20b"
 
-register_cuda_ci(est_time=100, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=142, suite="stage-b-test-1-gpu-large")
 
 
 class TestSWARadixCacheKL(KLDivergenceMixin, DefaultServerBase):
diff --git a/test/registered/reasoning/test_reasoning.py b/test/registered/reasoning/test_reasoning.py
index 0517e2c8f..b6295c3da 100644
--- a/test/registered/reasoning/test_reasoning.py
+++ b/test/registered/reasoning/test_reasoning.py
@@ -17,7 +17,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=109, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=111, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=200, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/rl/test_fp32_lm_head.py b/test/registered/rl/test_fp32_lm_head.py
index 8b721cc4e..f98769cdb 100644
--- a/test/registered/rl/test_fp32_lm_head.py
+++ b/test/registered/rl/test_fp32_lm_head.py
@@ -15,7 +15,7 @@ from sglang.srt.server_args import (
 from sglang.srt.utils import get_device
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=9, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=15, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/rl/test_lora_load_from_tensor.py b/test/registered/rl/test_lora_load_from_tensor.py
index 690eb266a..14bb64c7d 100644
--- a/test/registered/rl/test_lora_load_from_tensor.py
+++ b/test/registered/rl/test_lora_load_from_tensor.py
@@ -10,7 +10,7 @@ import sglang as sgl
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=90, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=78, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=90, suite="stage-b-test-1-gpu-small-amd")
 
 MODEL_PATH = "Qwen/Qwen3-0.6B"
diff --git a/test/registered/rl/test_multi_instance_release_memory_occupation.py b/test/registered/rl/test_multi_instance_release_memory_occupation.py
index c3f8d4b1d..5484b55b8 100644
--- a/test/registered/rl/test_multi_instance_release_memory_occupation.py
+++ b/test/registered/rl/test_multi_instance_release_memory_occupation.py
@@ -20,7 +20,7 @@ from sglang.test.test_utils import (
     find_available_port,
 )
 
-register_cuda_ci(est_time=64, suite="stage-c-test-4-gpu-h100")
+register_cuda_ci(est_time=57, suite="stage-c-test-4-gpu-h100")
 register_amd_ci(
     est_time=64,
     suite="stage-c-test-4-gpu-amd",
diff --git a/test/registered/rl/test_patch_torch.py b/test/registered/rl/test_patch_torch.py
index 2c46e6754..445713165 100644
--- a/test/registered/rl/test_patch_torch.py
+++ b/test/registered/rl/test_patch_torch.py
@@ -12,7 +12,7 @@ from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 register_amd_ci(
     est_time=19, suite="stage-b-test-2-gpu-large-amd", disabled="see #11127"
 )
-register_cuda_ci(est_time=19, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=22, suite="stage-b-test-2-gpu-large")
 
 
 class TestReleaseMemoryOccupation(unittest.TestCase):
diff --git a/test/registered/rl/test_pause_generation_tensor_consistency.py b/test/registered/rl/test_pause_generation_tensor_consistency.py
index 761a963e2..0223cca11 100644
--- a/test/registered/rl/test_pause_generation_tensor_consistency.py
+++ b/test/registered/rl/test_pause_generation_tensor_consistency.py
@@ -9,7 +9,7 @@ import torch
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=6, suite="stage-a-test-cpu")
 
 
 # ---------------------------------------------------------------------------
diff --git a/test/registered/rl/test_return_routed_experts.py b/test/registered/rl/test_return_routed_experts.py
index 7e3e3b517..d75bbd4e2 100644
--- a/test/registered/rl/test_return_routed_experts.py
+++ b/test/registered/rl/test_return_routed_experts.py
@@ -22,7 +22,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=200, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=164, suite="stage-b-test-2-gpu-large")
 register_amd_ci(
     est_time=200,
     suite="stage-b-test-2-gpu-large-amd",
diff --git a/test/registered/rl/test_update_weights_from_disk_mxfp8.py b/test/registered/rl/test_update_weights_from_disk_mxfp8.py
index a1111c138..12ac0758c 100644
--- a/test/registered/rl/test_update_weights_from_disk_mxfp8.py
+++ b/test/registered/rl/test_update_weights_from_disk_mxfp8.py
@@ -1,6 +1,6 @@
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=100, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=149, suite="stage-c-test-4-gpu-b200")
 
 import unittest
 
diff --git a/test/registered/rl/test_update_weights_from_distributed.py b/test/registered/rl/test_update_weights_from_distributed.py
index d0cad0285..3d464db3f 100644
--- a/test/registered/rl/test_update_weights_from_distributed.py
+++ b/test/registered/rl/test_update_weights_from_distributed.py
@@ -43,7 +43,7 @@ from sglang.test.test_utils import (
 )
 from sglang.utils import terminate_process
 
-register_cuda_ci(est_time=103, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=91, suite="stage-b-test-2-gpu-large")
 register_amd_ci(est_time=103, suite="stage-b-test-2-gpu-large-amd")
 
 mp.set_start_method("spawn", force=True)
diff --git a/test/registered/rl/test_update_weights_from_tensor.py b/test/registered/rl/test_update_weights_from_tensor.py
index aebe71e2a..cdd1afdc4 100644
--- a/test/registered/rl/test_update_weights_from_tensor.py
+++ b/test/registered/rl/test_update_weights_from_tensor.py
@@ -1,6 +1,6 @@
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=195, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=136, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=195, suite="stage-b-test-1-gpu-small-amd")
 
 import gc
diff --git a/test/registered/rotary/test_mrope.py b/test/registered/rotary/test_mrope.py
index 64a0b017b..87b34df91 100644
--- a/test/registered/rotary/test_mrope.py
+++ b/test/registered/rotary/test_mrope.py
@@ -20,7 +20,7 @@ from sglang.srt.utils import (
 )
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=10, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=7, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=15, suite="stage-b-test-1-gpu-small-amd")
 
 _is_cuda = is_cuda()
diff --git a/test/registered/sampling/test_original_logprobs.py b/test/registered/sampling/test_original_logprobs.py
index 47b4a9c57..b8125e407 100644
--- a/test/registered/sampling/test_original_logprobs.py
+++ b/test/registered/sampling/test_original_logprobs.py
@@ -25,7 +25,7 @@ import sglang as sgl
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import DEFAULT_SMALL_MODEL_NAME_FOR_TEST
 
-register_cuda_ci(est_time=41, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=46, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=60, suite="stage-b-test-1-gpu-small-amd")
 
 # ------------------------- Configurable via env ------------------------- #
diff --git a/test/registered/sampling/test_penalty.py b/test/registered/sampling/test_penalty.py
index df456e0f7..06311c083 100644
--- a/test/registered/sampling/test_penalty.py
+++ b/test/registered/sampling/test_penalty.py
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=82, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=50, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=82, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/sampling/test_pytorch_sampling_backend.py b/test/registered/sampling/test_pytorch_sampling_backend.py
index 3b6df45f9..e53cb1df5 100644
--- a/test/registered/sampling/test_pytorch_sampling_backend.py
+++ b/test/registered/sampling/test_pytorch_sampling_backend.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=80, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=77, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=66, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/scheduler/test_abort.py b/test/registered/scheduler/test_abort.py
index 47399bc4e..b82a22e9a 100644
--- a/test/registered/scheduler/test_abort.py
+++ b/test/registered/scheduler/test_abort.py
@@ -19,7 +19,7 @@ from sglang.test.test_utils import (
     run_and_check_memory_leak,
 )
 
-register_cuda_ci(est_time=350, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=328, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=300, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/scheduler/test_no_chunked_prefill.py b/test/registered/scheduler/test_no_chunked_prefill.py
index a255346d0..5aa67fc14 100644
--- a/test/registered/scheduler/test_no_chunked_prefill.py
+++ b/test/registered/scheduler/test_no_chunked_prefill.py
@@ -8,7 +8,7 @@ from sglang.test.test_utils import (
     run_mmlu_test,
 )
 
-register_cuda_ci(est_time=108, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=121, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=108, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/scheduler/test_no_overlap_scheduler.py b/test/registered/scheduler/test_no_overlap_scheduler.py
index 8ae06fe16..b3e0b28f6 100644
--- a/test/registered/scheduler/test_no_overlap_scheduler.py
+++ b/test/registered/scheduler/test_no_overlap_scheduler.py
@@ -9,7 +9,7 @@ import unittest
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase, run_mmlu_test
 
-register_cuda_ci(est_time=245, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=231, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=275, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/scheduler/test_priority_scheduling.py b/test/registered/scheduler/test_priority_scheduling.py
index 72321b694..c1244596c 100644
--- a/test/registered/scheduler/test_priority_scheduling.py
+++ b/test/registered/scheduler/test_priority_scheduling.py
@@ -17,7 +17,7 @@ from sglang.test.test_utils import (
     send_concurrent_generate_requests_with_custom_params,
 )
 
-register_cuda_ci(est_time=130, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=133, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=195, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/scheduler/test_retract_decode.py b/test/registered/scheduler/test_retract_decode.py
index 42735b245..45e120ca1 100644
--- a/test/registered/scheduler/test_retract_decode.py
+++ b/test/registered/scheduler/test_retract_decode.py
@@ -17,7 +17,7 @@ from sglang.test.test_utils import (
 )
 from sglang.utils import is_in_ci
 
-register_cuda_ci(est_time=470, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=305, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=600, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/sessions/test_session_control.py b/test/registered/sessions/test_session_control.py
index 0e6b2a7f2..386d57b9b 100644
--- a/test/registered/sessions/test_session_control.py
+++ b/test/registered/sessions/test_session_control.py
@@ -24,7 +24,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=60, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=77, suite="stage-b-test-1-gpu-large")
 
 
 def remove_prefix(text: str, prefix: str) -> str:
diff --git a/test/registered/sessions/test_streaming_session.py b/test/registered/sessions/test_streaming_session.py
index ed2da9cc3..ffefe30be 100644
--- a/test/registered/sessions/test_streaming_session.py
+++ b/test/registered/sessions/test_streaming_session.py
@@ -28,7 +28,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=70, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=67, suite="stage-b-test-1-gpu-large")
 
 # ---------------------------------------------------------------------------
 # Logprob leak constants
diff --git a/test/registered/spec/dflash/test_dflash.py b/test/registered/spec/dflash/test_dflash.py
index aa9ee2327..1583bd92b 100644
--- a/test/registered/spec/dflash/test_dflash.py
+++ b/test/registered/spec/dflash/test_dflash.py
@@ -18,7 +18,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=300, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=270, suite="stage-b-test-1-gpu-small")
 
 
 class TestDFlashServerBase(CustomTestCase, MatchedStopMixin, GSM8KMixin):
diff --git a/test/registered/spec/eagle/test_deepseek_v3_fp4_mtp_small.py b/test/registered/spec/eagle/test_deepseek_v3_fp4_mtp_small.py
index fe589679f..7b92a5803 100644
--- a/test/registered/spec/eagle/test_deepseek_v3_fp4_mtp_small.py
+++ b/test/registered/spec/eagle/test_deepseek_v3_fp4_mtp_small.py
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=240, suite="stage-b-test-4-gpu-b200")
+register_cuda_ci(est_time=416, suite="stage-b-test-4-gpu-b200")
 
 FULL_DEEPSEEK_V3_FP4_MODEL_PATH = "nvidia/DeepSeek-V3-0324-FP4"
 SERVER_LAUNCH_TIMEOUT = 1200
diff --git a/test/registered/spec/eagle/test_eagle3_basic.py b/test/registered/spec/eagle/test_eagle3_basic.py
index 2830228a3..19486829e 100644
--- a/test/registered/spec/eagle/test_eagle3_basic.py
+++ b/test/registered/spec/eagle/test_eagle3_basic.py
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     DEFAULT_TARGET_MODEL_EAGLE3,
 )
 
-register_cuda_ci(est_time=160, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=70, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=50, suite="stage-b-test-1-gpu-small")
 
 _is_hip = is_hip()
diff --git a/test/registered/spec/eagle/test_eagle_constrained_decoding.py b/test/registered/spec/eagle/test_eagle_constrained_decoding.py
index 6fa1dd5a3..13712d152 100644
--- a/test/registered/spec/eagle/test_eagle_constrained_decoding.py
+++ b/test/registered/spec/eagle/test_eagle_constrained_decoding.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=100, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=96, suite="stage-b-test-1-gpu-large")
 
 
 class TestEagleConstrainedDecoding(
diff --git a/test/registered/spec/eagle/test_eagle_dp_attention.py b/test/registered/spec/eagle/test_eagle_dp_attention.py
index fe04bdb10..3f9dc9d1c 100644
--- a/test/registered/spec/eagle/test_eagle_dp_attention.py
+++ b/test/registered/spec/eagle/test_eagle_dp_attention.py
@@ -21,7 +21,7 @@ from sglang.test.test_utils import (
 )
 
 # EAGLE3 with DP attention (tp=2, dp=2, requires 4 GPUs)
-register_cuda_ci(est_time=100, suite="stage-c-test-4-gpu-h100")
+register_cuda_ci(est_time=96, suite="stage-c-test-4-gpu-h100")
 register_amd_ci(est_time=200, suite="stage-c-test-4-gpu-amd")
 
 
diff --git a/test/registered/spec/eagle/test_eagle_infer_a.py b/test/registered/spec/eagle/test_eagle_infer_a.py
index bce64a0c4..e7bfd9936 100644
--- a/test/registered/spec/eagle/test_eagle_infer_a.py
+++ b/test/registered/spec/eagle/test_eagle_infer_a.py
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     CustomTestCase,
 )
 
-register_cuda_ci(est_time=270, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=266, suite="stage-b-test-1-gpu-large")
 
 
 class TestEAGLEEngine(CustomTestCase):
diff --git a/test/registered/spec/eagle/test_eagle_infer_b.py b/test/registered/spec/eagle/test_eagle_infer_b.py
index 1585a4bcc..c603a04d8 100644
--- a/test/registered/spec/eagle/test_eagle_infer_b.py
+++ b/test/registered/spec/eagle/test_eagle_infer_b.py
@@ -22,7 +22,7 @@ from sglang.test.run_eval import run_eval
 from sglang.test.server_fixtures.eagle_fixture import EagleServerBase
 from sglang.test.test_utils import DEFAULT_TARGET_MODEL_EAGLE, run_logprob_check
 
-register_cuda_ci(est_time=690, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=684, suite="stage-b-test-1-gpu-large")
 
 
 class TestEAGLEServerBasic(EagleServerBase):
diff --git a/test/registered/spec/eagle/test_eagle_infer_beta.py b/test/registered/spec/eagle/test_eagle_infer_beta.py
index efaa9d6d8..77a571c14 100644
--- a/test/registered/spec/eagle/test_eagle_infer_beta.py
+++ b/test/registered/spec/eagle/test_eagle_infer_beta.py
@@ -19,7 +19,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=283, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=333, suite="stage-b-test-1-gpu-small")
 
 
 class TestEagle3ServerBase(CustomTestCase, MatchedStopMixin):
diff --git a/test/registered/spec/eagle/test_eagle_infer_beta_dp_attention.py b/test/registered/spec/eagle/test_eagle_infer_beta_dp_attention.py
index 407004dc0..0b3d75f60 100644
--- a/test/registered/spec/eagle/test_eagle_infer_beta_dp_attention.py
+++ b/test/registered/spec/eagle/test_eagle_infer_beta_dp_attention.py
@@ -17,7 +17,7 @@ from sglang.test.test_utils import (
 )
 
 # EAGLE with DP attention on B200 (tp=2, dp=2, requires 4 B200 GPUs)
-register_cuda_ci(est_time=100, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=68, suite="stage-c-test-4-gpu-b200")
 
 
 def test_gsm8k(base_url: str, model: str):
diff --git a/test/registered/spec/test_constrained_decoding_spec_reasoning.py b/test/registered/spec/test_constrained_decoding_spec_reasoning.py
index 07f59bd2d..81e53d01e 100644
--- a/test/registered/spec/test_constrained_decoding_spec_reasoning.py
+++ b/test/registered/spec/test_constrained_decoding_spec_reasoning.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
 )
 
 # Constrained decoding with EAGLE3 speculative reasoning (tp=2)
-register_cuda_ci(est_time=60, suite="stage-b-test-2-gpu-large")
+register_cuda_ci(est_time=80, suite="stage-b-test-2-gpu-large")
 
 
 class ServerWithGrammar(CustomTestCase):
diff --git a/test/registered/spec/test_ngram_speculative_decoding.py b/test/registered/spec/test_ngram_speculative_decoding.py
index d8e0c467b..f1fdfb11c 100644
--- a/test/registered/spec/test_ngram_speculative_decoding.py
+++ b/test/registered/spec/test_ngram_speculative_decoding.py
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=230, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=200, suite="stage-b-test-1-gpu-large")
 
 GSM_DATASET_PATH = None
 
diff --git a/test/registered/spec/test_standalone_speculative_decoding.py b/test/registered/spec/test_standalone_speculative_decoding.py
index 37cb6fae7..cebe13467 100644
--- a/test/registered/spec/test_standalone_speculative_decoding.py
+++ b/test/registered/spec/test_standalone_speculative_decoding.py
@@ -18,7 +18,7 @@ from sglang.test.test_utils import (
 )
 
 # Standalone speculative decoding tests (FA3, Triton, FlashInfer backends)
-register_cuda_ci(est_time=308, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=294, suite="stage-b-test-1-gpu-large")
 
 GSM_DATASET_PATH = None
 
diff --git a/test/registered/spec/utils/test_build_eagle_tree.py b/test/registered/spec/utils/test_build_eagle_tree.py
index 103349fcc..d07afeb7c 100644
--- a/test/registered/spec/utils/test_build_eagle_tree.py
+++ b/test/registered/spec/utils/test_build_eagle_tree.py
@@ -9,7 +9,7 @@ from sglang.srt.speculative.eagle_utils import (
 from sglang.srt.utils import get_device
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=6, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=4, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=3, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/tokenizer/test_multi_tokenizer.py b/test/registered/tokenizer/test_multi_tokenizer.py
index b367cd4f8..3fb763159 100644
--- a/test/registered/tokenizer/test_multi_tokenizer.py
+++ b/test/registered/tokenizer/test_multi_tokenizer.py
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=230, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=204, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=345, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/tokenizer/test_skip_tokenizer_init.py b/test/registered/tokenizer/test_skip_tokenizer_init.py
index 5f821030f..3cb491979 100644
--- a/test/registered/tokenizer/test_skip_tokenizer_init.py
+++ b/test/registered/tokenizer/test_skip_tokenizer_init.py
@@ -23,7 +23,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=77, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=82, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=117, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/unit/auto_benchmark/test_dataset_tools.py b/test/registered/unit/auto_benchmark/test_dataset_tools.py
index b302ad86f..821921289 100644
--- a/test/registered/unit/auto_benchmark/test_dataset_tools.py
+++ b/test/registered/unit/auto_benchmark/test_dataset_tools.py
@@ -15,7 +15,7 @@ from sglang.auto_benchmark_lib import infer_backend, prepare_dataset
 from sglang.benchmark.datasets.autobench import sample_autobench_requests
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=1, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=5, suite="stage-b-test-1-gpu-small")
 
 
 class TestAutoBenchmarkDatasetTools(AutoBenchmarkTestCase):
diff --git a/test/registered/unit/auto_benchmark/test_run_candidate.py b/test/registered/unit/auto_benchmark/test_run_candidate.py
index ee919f2c1..7c3b78f22 100644
--- a/test/registered/unit/auto_benchmark/test_run_candidate.py
+++ b/test/registered/unit/auto_benchmark/test_run_candidate.py
@@ -14,7 +14,7 @@ from auto_benchmark import AutoBenchmarkTestCase
 from sglang.auto_benchmark_lib import SearchDeadlineExceeded, run_candidate
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=2, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=5, suite="stage-b-test-1-gpu-small")
 
 
 class TestAutoBenchmarkRunCandidate(AutoBenchmarkTestCase):
diff --git a/test/registered/unit/auto_benchmark/test_search_tools.py b/test/registered/unit/auto_benchmark/test_search_tools.py
index 2f304bca7..409d7e2f7 100644
--- a/test/registered/unit/auto_benchmark/test_search_tools.py
+++ b/test/registered/unit/auto_benchmark/test_search_tools.py
@@ -28,7 +28,7 @@ from sglang.auto_benchmark_lib import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=2, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=5, suite="stage-b-test-1-gpu-small")
 
 
 class TestAutoBenchmarkSearchTools(AutoBenchmarkTestCase):
diff --git a/test/registered/unit/configs/test_linear_attn_model_registry.py b/test/registered/unit/configs/test_linear_attn_model_registry.py
index 6fbd27d0a..8933a25d5 100644
--- a/test/registered/unit/configs/test_linear_attn_model_registry.py
+++ b/test/registered/unit/configs/test_linear_attn_model_registry.py
@@ -13,7 +13,7 @@ from sglang.srt.configs.linear_attn_model_registry import (
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=6, suite="stage-a-test-cpu")
 
 
 # Dummy config classes for testing
diff --git a/test/registered/unit/entrypoints/openai/test_serving_embedding.py b/test/registered/unit/entrypoints/openai/test_serving_embedding.py
index abbec63ec..ccbd03435 100644
--- a/test/registered/unit/entrypoints/openai/test_serving_embedding.py
+++ b/test/registered/unit/entrypoints/openai/test_serving_embedding.py
@@ -54,7 +54,7 @@ from sglang.srt.entrypoints.openai.serving_embedding import OpenAIServingEmbeddi
 from sglang.srt.managers.io_struct import EmbeddingReqInput
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=10, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=9, suite="stage-a-test-cpu")
 
 
 # Mock TokenizerManager for embedding tests
diff --git a/test/registered/unit/entrypoints/test_ssl_cert_refresher.py b/test/registered/unit/entrypoints/test_ssl_cert_refresher.py
index 8b8283d06..85393fd86 100644
--- a/test/registered/unit/entrypoints/test_ssl_cert_refresher.py
+++ b/test/registered/unit/entrypoints/test_ssl_cert_refresher.py
@@ -8,7 +8,7 @@ from sglang.srt.entrypoints.ssl_utils import SSLCertRefresher
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cpu_ci(est_time=9, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=13, suite="stage-a-test-cpu")
 
 
 def _make_temp_pem(content: bytes) -> str:
diff --git a/test/registered/unit/layers/test_conv_layer.py b/test/registered/unit/layers/test_conv_layer.py
index b2520f4d2..20ab6698a 100644
--- a/test/registered/unit/layers/test_conv_layer.py
+++ b/test/registered/unit/layers/test_conv_layer.py
@@ -1,6 +1,6 @@
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=10, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=4, suite="stage-b-test-1-gpu-small")
 
 import unittest
 
diff --git a/test/registered/unit/layers/test_mamba_state_scatter_triton.py b/test/registered/unit/layers/test_mamba_state_scatter_triton.py
index 1a94dc6aa..c1f106c4b 100644
--- a/test/registered/unit/layers/test_mamba_state_scatter_triton.py
+++ b/test/registered/unit/layers/test_mamba_state_scatter_triton.py
@@ -1,6 +1,6 @@
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=10, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=4, suite="stage-b-test-1-gpu-small")
 
 import os
 import unittest
diff --git a/test/registered/unit/layers/test_pooler_score_and_pool.py b/test/registered/unit/layers/test_pooler_score_and_pool.py
index 374e7ff02..b40c5afca 100644
--- a/test/registered/unit/layers/test_pooler_score_and_pool.py
+++ b/test/registered/unit/layers/test_pooler_score_and_pool.py
@@ -20,7 +20,7 @@ from sglang.srt.layers.pooler import (
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=9, suite="stage-a-test-cpu")
 
 
 def _make_forward_batch(extend_seq_lens, is_prefill_only=False):
diff --git a/test/registered/unit/managers/test_profile_merger_http_api.py b/test/registered/unit/managers/test_profile_merger_http_api.py
index 1270fdf83..53e49feb4 100644
--- a/test/registered/unit/managers/test_profile_merger_http_api.py
+++ b/test/registered/unit/managers/test_profile_merger_http_api.py
@@ -4,7 +4,7 @@ import unittest
 from sglang.srt.managers.io_struct import ProfileReqInput
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=9, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=9, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/unit/managers/test_scheduler_flush_cache.py b/test/registered/unit/managers/test_scheduler_flush_cache.py
index 828854ed4..249b4b8ce 100644
--- a/test/registered/unit/managers/test_scheduler_flush_cache.py
+++ b/test/registered/unit/managers/test_scheduler_flush_cache.py
@@ -9,7 +9,7 @@ maybe_stub_sgl_kernel()
 from sglang.srt.managers.io_struct import FlushCacheReqInput
 from sglang.srt.managers.scheduler import Scheduler
 
-register_cpu_ci(est_time=2, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=10, suite="stage-a-test-cpu")
 
 
 class TestSchedulerFlushCache(unittest.TestCase):
diff --git a/test/registered/unit/managers/test_scheduler_pause_generation.py b/test/registered/unit/managers/test_scheduler_pause_generation.py
index 210ba0aa6..4e94626d4 100644
--- a/test/registered/unit/managers/test_scheduler_pause_generation.py
+++ b/test/registered/unit/managers/test_scheduler_pause_generation.py
@@ -10,7 +10,7 @@ maybe_stub_sgl_kernel()
 from sglang.srt.managers.io_struct import PauseGenerationReqInput
 from sglang.srt.managers.scheduler import Scheduler
 
-register_cpu_ci(est_time=2, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=10, suite="stage-a-test-cpu")
 
 
 class TestSchedulerPauseGeneration(unittest.TestCase):
diff --git a/test/registered/unit/mem_cache/test_evict_policy.py b/test/registered/unit/mem_cache/test_evict_policy.py
index c59af743f..13e022acf 100644
--- a/test/registered/unit/mem_cache/test_evict_policy.py
+++ b/test/registered/unit/mem_cache/test_evict_policy.py
@@ -2,7 +2,7 @@
 
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=2, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=1, suite="stage-a-test-cpu")
 
 import unittest
 from unittest.mock import MagicMock
diff --git a/test/registered/unit/mem_cache/test_mamba_unittest.py b/test/registered/unit/mem_cache/test_mamba_unittest.py
index c6219a46d..adadc5f3c 100644
--- a/test/registered/unit/mem_cache/test_mamba_unittest.py
+++ b/test/registered/unit/mem_cache/test_mamba_unittest.py
@@ -21,7 +21,7 @@ from sglang.srt.server_args import ServerArgs, set_global_server_args_for_schedu
 from sglang.srt.utils import get_device
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=9, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=9, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/unit/mem_cache/test_nsa_pool_host_unit.py b/test/registered/unit/mem_cache/test_nsa_pool_host_unit.py
index 64ca27493..f0462e267 100644
--- a/test/registered/unit/mem_cache/test_nsa_pool_host_unit.py
+++ b/test/registered/unit/mem_cache/test_nsa_pool_host_unit.py
@@ -11,7 +11,7 @@ from sglang.srt.mem_cache.memory_pool_host import (
 from sglang.srt.utils import is_cuda, is_hip, is_npu, is_xpu
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=10, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-small")
 
 
 class TestNSAHiCacheTransfer(unittest.TestCase):
diff --git a/test/registered/unit/mem_cache/test_radix_cache_slru_accuracy.py b/test/registered/unit/mem_cache/test_radix_cache_slru_accuracy.py
index 7d496f2e9..e407c1741 100644
--- a/test/registered/unit/mem_cache/test_radix_cache_slru_accuracy.py
+++ b/test/registered/unit/mem_cache/test_radix_cache_slru_accuracy.py
@@ -13,7 +13,7 @@ from sglang.srt.mem_cache.memory_pool import MHATokenToKVPool, ReqToTokenPool
 from sglang.srt.mem_cache.radix_cache import RadixCache, RadixKey
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=5, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=7, suite="stage-b-test-1-gpu-small")
 
 
 class TestSLRUAccuracy(unittest.TestCase):
diff --git a/test/registered/unit/mem_cache/test_radix_cache_unit.py b/test/registered/unit/mem_cache/test_radix_cache_unit.py
index f7b8a43ce..f01bc7736 100644
--- a/test/registered/unit/mem_cache/test_radix_cache_unit.py
+++ b/test/registered/unit/mem_cache/test_radix_cache_unit.py
@@ -21,7 +21,7 @@ from sglang.srt.mem_cache.common import available_and_evictable_str
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
 # CPU-based unit test, runs quickly on any GPU runner
-register_cuda_ci(est_time=20, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=16, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=5, suite="stage-b-test-1-gpu-small-amd")
 
 import random
diff --git a/test/registered/unit/model_executor/test_pool_configurator.py b/test/registered/unit/model_executor/test_pool_configurator.py
index e196045f9..04dc59e12 100644
--- a/test/registered/unit/model_executor/test_pool_configurator.py
+++ b/test/registered/unit/model_executor/test_pool_configurator.py
@@ -12,7 +12,7 @@ from unittest.mock import MagicMock, patch
 
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=9, suite="stage-a-test-cpu")
 
 
 @contextlib.contextmanager
diff --git a/test/registered/unit/model_loader/test_modelopt_export.py b/test/registered/unit/model_loader/test_modelopt_export.py
index be448d467..8c2d634de 100644
--- a/test/registered/unit/model_loader/test_modelopt_export.py
+++ b/test/registered/unit/model_loader/test_modelopt_export.py
@@ -19,7 +19,7 @@ from sglang.srt.configs.model_config import ModelConfig
 from sglang.srt.model_loader.loader import ModelOptModelLoader
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=9, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=9, suite="stage-b-test-1-gpu-small-amd")
 
 # Note: PYTHONPATH=python should be set when running tests
diff --git a/test/registered/unit/model_loader/test_modelopt_loader.py b/test/registered/unit/model_loader/test_modelopt_loader.py
index 7f9652c0e..a7afcaf24 100644
--- a/test/registered/unit/model_loader/test_modelopt_loader.py
+++ b/test/registered/unit/model_loader/test_modelopt_loader.py
@@ -30,7 +30,7 @@ CALIBRATION_BATCH_SIZE = 36
 CALIBRATION_NUM_SAMPLES = 512
 DEFAULT_DEVICE = "cuda:0"
 
-register_cuda_ci(est_time=11, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=10, suite="stage-b-test-1-gpu-small")
 
 
 class TestModelOptModelLoader(CustomTestCase):
diff --git a/test/registered/unit/models/test_llava.py b/test/registered/unit/models/test_llava.py
index a929dcfbd..1e4ca1ce2 100644
--- a/test/registered/unit/models/test_llava.py
+++ b/test/registered/unit/models/test_llava.py
@@ -5,7 +5,7 @@ from sglang.srt.models.llava import AutoModel, LlavaForConditionalGeneration
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=1, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-small")
 
 
 class PixtralVisionConfig:
diff --git a/test/registered/unit/observability/test_func_timer.py b/test/registered/unit/observability/test_func_timer.py
index 397fdafb7..988a5ca27 100644
--- a/test/registered/unit/observability/test_func_timer.py
+++ b/test/registered/unit/observability/test_func_timer.py
@@ -2,7 +2,7 @@
 
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=1, suite="stage-a-test-cpu")
 
 import asyncio
 import unittest
diff --git a/test/registered/unit/observability/test_request_metrics_exporter.py b/test/registered/unit/observability/test_request_metrics_exporter.py
index 921c5dd1d..3dd4f618c 100644
--- a/test/registered/unit/observability/test_request_metrics_exporter.py
+++ b/test/registered/unit/observability/test_request_metrics_exporter.py
@@ -5,7 +5,7 @@ from typing import Any, Dict, List, Optional
 
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=1, suite="stage-a-test-cpu")
 
 import asyncio
 import json
diff --git a/test/registered/unit/observability/test_startup_func_log_and_timer.py b/test/registered/unit/observability/test_startup_func_log_and_timer.py
index 47281cf2d..5afcfab8f 100644
--- a/test/registered/unit/observability/test_startup_func_log_and_timer.py
+++ b/test/registered/unit/observability/test_startup_func_log_and_timer.py
@@ -2,7 +2,7 @@
 
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=1, suite="stage-a-test-cpu")
 
 import unittest
 from unittest.mock import MagicMock, patch
diff --git a/test/registered/unit/observability/test_trace.py b/test/registered/unit/observability/test_trace.py
index 4a5efc92e..78654030c 100644
--- a/test/registered/unit/observability/test_trace.py
+++ b/test/registered/unit/observability/test_trace.py
@@ -4,7 +4,7 @@ import os
 
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=3, suite="stage-a-test-cpu")
 
 import threading
 import unittest
diff --git a/test/registered/unit/parser/test_code_completion_parser.py b/test/registered/unit/parser/test_code_completion_parser.py
index 1ee5f1f94..1182e0910 100644
--- a/test/registered/unit/parser/test_code_completion_parser.py
+++ b/test/registered/unit/parser/test_code_completion_parser.py
@@ -18,7 +18,7 @@ from sglang.srt.parser.code_completion_parser import (
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=6, suite="stage-a-test-cpu")
 
 
 class TestFimPosition(CustomTestCase):
diff --git a/test/registered/unit/parser/test_conversation.py b/test/registered/unit/parser/test_conversation.py
index b42445068..ed0a43c44 100644
--- a/test/registered/unit/parser/test_conversation.py
+++ b/test/registered/unit/parser/test_conversation.py
@@ -32,7 +32,7 @@ from sglang.srt.parser.conversation import (
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=6, suite="stage-a-test-cpu")
 
 
 class TestConversationGetPrompt(CustomTestCase):
diff --git a/test/registered/unit/parser/test_jinja_template_utils.py b/test/registered/unit/parser/test_jinja_template_utils.py
index 3396de98a..c3845148e 100644
--- a/test/registered/unit/parser/test_jinja_template_utils.py
+++ b/test/registered/unit/parser/test_jinja_template_utils.py
@@ -9,7 +9,7 @@ from sglang.srt.parser.jinja_template_utils import (
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cpu_ci(est_time=7, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=6, suite="stage-a-test-cpu")
 
 
 class TestTemplateContentFormatDetection(CustomTestCase):
diff --git a/test/registered/unit/parser/test_reasoning_content_without_parser.py b/test/registered/unit/parser/test_reasoning_content_without_parser.py
index 446dea6b3..265adb711 100644
--- a/test/registered/unit/parser/test_reasoning_content_without_parser.py
+++ b/test/registered/unit/parser/test_reasoning_content_without_parser.py
@@ -4,7 +4,7 @@ from sglang.srt.parser.reasoning_parser import ReasoningParser
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=6, suite="stage-a-test-cpu")
 
 # Simulated model output that contains think tags (e.g. from DeepSeek-R1)
 THINK_OUTPUT = (
diff --git a/test/registered/unit/parser/test_reasoning_parser.py b/test/registered/unit/parser/test_reasoning_parser.py
index 8f05d7903..92ad536a8 100644
--- a/test/registered/unit/parser/test_reasoning_parser.py
+++ b/test/registered/unit/parser/test_reasoning_parser.py
@@ -17,7 +17,7 @@ from sglang.srt.parser.reasoning_parser import (
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=6, suite="stage-a-test-cpu")
 
 
 class TestStreamingParseResult(CustomTestCase):
diff --git a/test/registered/unit/sampling/test_custom_logit_processor.py b/test/registered/unit/sampling/test_custom_logit_processor.py
index 00922b45b..6ef5544aa 100644
--- a/test/registered/unit/sampling/test_custom_logit_processor.py
+++ b/test/registered/unit/sampling/test_custom_logit_processor.py
@@ -2,7 +2,7 @@
 
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=6, suite="stage-a-test-cpu")
 
 import json
 import unittest
diff --git a/test/registered/unit/sampling/test_penaltylib.py b/test/registered/unit/sampling/test_penaltylib.py
index 8681bcb34..4943ff887 100644
--- a/test/registered/unit/sampling/test_penaltylib.py
+++ b/test/registered/unit/sampling/test_penaltylib.py
@@ -2,7 +2,7 @@
 
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=8, suite="stage-a-test-cpu")
 
 import unittest
 from unittest.mock import MagicMock
diff --git a/test/registered/unit/sampling/test_sampling_batch_info.py b/test/registered/unit/sampling/test_sampling_batch_info.py
index 7b0189153..9af7587f2 100644
--- a/test/registered/unit/sampling/test_sampling_batch_info.py
+++ b/test/registered/unit/sampling/test_sampling_batch_info.py
@@ -2,7 +2,7 @@
 
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=8, suite="stage-a-test-cpu")
 
 import unittest
 from unittest.mock import MagicMock, patch
diff --git a/test/registered/unit/sampling/test_sampling_params.py b/test/registered/unit/sampling/test_sampling_params.py
index 4d2431bf9..ce9a7c3c3 100644
--- a/test/registered/unit/sampling/test_sampling_params.py
+++ b/test/registered/unit/sampling/test_sampling_params.py
@@ -2,7 +2,7 @@
 
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=6, suite="stage-a-test-cpu")
 
 import unittest
 from unittest.mock import MagicMock
diff --git a/test/registered/unit/server_args/test_server_args.py b/test/registered/unit/server_args/test_server_args.py
index a83a2c693..29b962b74 100644
--- a/test/registered/unit/server_args/test_server_args.py
+++ b/test/registered/unit/server_args/test_server_args.py
@@ -10,7 +10,7 @@ from sglang.test.test_utils import (
     CustomTestCase,
 )
 
-register_cpu_ci(est_time=10, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=9, suite="stage-a-test-cpu")
 
 # Mock get_device() so all tests run on CPU-only CI runners
 _mock_device = patch("sglang.srt.server_args.get_device", return_value="cuda")
diff --git a/test/registered/unit/spec/test_ngram_corpus.py b/test/registered/unit/spec/test_ngram_corpus.py
index 62ae418e3..1ecc30b4a 100644
--- a/test/registered/unit/spec/test_ngram_corpus.py
+++ b/test/registered/unit/spec/test_ngram_corpus.py
@@ -13,7 +13,7 @@ from sglang.srt.speculative.cpp_ngram.ngram_corpus import NgramCorpus
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cpu_ci(est_time=10, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=18, suite="stage-a-test-cpu")
 
 
 def _make_corpus(match_type="BFS", **kwargs):
diff --git a/test/registered/unit/test_runai_utils.py b/test/registered/unit/test_runai_utils.py
index 4735b5d7f..865b3c7a9 100644
--- a/test/registered/unit/test_runai_utils.py
+++ b/test/registered/unit/test_runai_utils.py
@@ -6,7 +6,7 @@ from sglang.srt.utils.runai_utils import ObjectStorageModel, is_runai_obj_uri
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=6, suite="stage-a-test-cpu")
 
 
 class TestRunaiUtils(CustomTestCase):
diff --git a/test/registered/unit/tokenizer/test_tiktoken_tokenizer.py b/test/registered/unit/tokenizer/test_tiktoken_tokenizer.py
index 7a6f4d9dc..c683d0178 100644
--- a/test/registered/unit/tokenizer/test_tiktoken_tokenizer.py
+++ b/test/registered/unit/tokenizer/test_tiktoken_tokenizer.py
@@ -6,7 +6,7 @@ from unittest.mock import MagicMock, patch
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cpu_ci(est_time=1, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=6, suite="stage-a-test-cpu")
 
 from sglang.srt.tokenizer.tiktoken_tokenizer import (
     CONTROL_TOKEN_TEXTS,
diff --git a/test/registered/unit/utils/test_gauge_histogram.py b/test/registered/unit/utils/test_gauge_histogram.py
index 5f8dab542..e208d3e7a 100644
--- a/test/registered/unit/utils/test_gauge_histogram.py
+++ b/test/registered/unit/utils/test_gauge_histogram.py
@@ -1,6 +1,6 @@
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=3, suite="stage-a-test-cpu")
 
 import unittest
 
diff --git a/test/registered/unit/utils/test_http_server_auth.py b/test/registered/unit/utils/test_http_server_auth.py
index 226c3bf88..25c948704 100644
--- a/test/registered/unit/utils/test_http_server_auth.py
+++ b/test/registered/unit/utils/test_http_server_auth.py
@@ -10,7 +10,7 @@ import unittest
 from sglang.srt.utils.auth import AuthLevel, decide_request_auth
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=3, suite="stage-a-test-cpu")
 
 
 class TestHttpServerAdminAuth(unittest.TestCase):
diff --git a/test/registered/unit/utils/test_json_response.py b/test/registered/unit/utils/test_json_response.py
index accc71fc5..20669d4ac 100644
--- a/test/registered/unit/utils/test_json_response.py
+++ b/test/registered/unit/utils/test_json_response.py
@@ -10,7 +10,7 @@ from sglang.srt.utils.json_response import (
 )
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=4, suite="stage-a-test-cpu")
 
 
 class TestJSONResponseUtils(unittest.TestCase):
diff --git a/test/registered/unit/utils/test_subprocess_watchdog.py b/test/registered/unit/utils/test_subprocess_watchdog.py
index 2de995616..1e9f50f92 100644
--- a/test/registered/unit/utils/test_subprocess_watchdog.py
+++ b/test/registered/unit/utils/test_subprocess_watchdog.py
@@ -24,7 +24,7 @@ from sglang.srt.utils.watchdog import SubprocessWatchdog
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cpu_ci(est_time=10, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=8, suite="stage-a-test-cpu")
 
 
 def healthy_worker():
diff --git a/test/registered/utils/test_log_utils.py b/test/registered/utils/test_log_utils.py
index 74341c13b..449b8a856 100644
--- a/test/registered/utils/test_log_utils.py
+++ b/test/registered/utils/test_log_utils.py
@@ -9,7 +9,7 @@ from pathlib import Path
 from sglang.srt.utils.log_utils import create_log_targets, log_json
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=4, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=3, suite="stage-a-test-cpu")
 
 
 class TestLogUtils(unittest.TestCase):
diff --git a/test/registered/utils/test_network_address.py b/test/registered/utils/test_network_address.py
index ff05206c2..529162f2c 100644
--- a/test/registered/utils/test_network_address.py
+++ b/test/registered/utils/test_network_address.py
@@ -6,7 +6,7 @@ from sglang.srt.server_args import PortArgs, ServerArgs
 from sglang.srt.utils.network import NetworkAddress
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=7, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=6, suite="stage-a-test-cpu")
 
 # Mock get_device() so ServerArgs tests run on CPU-only CI runners
 _mock_device = patch("sglang.srt.server_args.get_device", return_value="cuda")
diff --git a/test/registered/utils/test_numa_utils.py b/test/registered/utils/test_numa_utils.py
index 1e084fef9..fc3fe61bc 100644
--- a/test/registered/utils/test_numa_utils.py
+++ b/test/registered/utils/test_numa_utils.py
@@ -9,7 +9,7 @@ from sglang.srt.utils.numa_utils import (
 )
 from sglang.test.ci.ci_register import register_cpu_ci, register_cuda_ci
 
-register_cpu_ci(est_time=1, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=6, suite="stage-a-test-cpu")
 register_cuda_ci(est_time=10, suite="stage-c-test-4-gpu-gb200")
 register_cuda_ci(est_time=10, suite="stage-c-test-8-gpu-b200")
 
diff --git a/test/registered/utils/test_socket_utils.py b/test/registered/utils/test_socket_utils.py
index 63a42ac39..058b81711 100644
--- a/test/registered/utils/test_socket_utils.py
+++ b/test/registered/utils/test_socket_utils.py
@@ -15,7 +15,7 @@ from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 from sglang.utils import normalize_base_url, release_port, reserve_port
 
-register_cpu_ci(est_time=7, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=6, suite="stage-a-test-cpu")
 
 
 class TestTryBindSocket(CustomTestCase):
diff --git a/test/registered/vlm/test_evs.py b/test/registered/vlm/test_evs.py
index 1f1da070f..cd4aa8b48 100644
--- a/test/registered/vlm/test_evs.py
+++ b/test/registered/vlm/test_evs.py
@@ -6,7 +6,7 @@ import pytest
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import run_doctests
 
-register_cuda_ci(est_time=20, suite="stage-b-test-1-gpu-small")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-small")
 register_amd_ci(est_time=20, suite="stage-b-test-1-gpu-small-amd")
 
 
diff --git a/test/registered/vlm/test_patch_embed_perf.py b/test/registered/vlm/test_patch_embed_perf.py
index 3ef3a6457..052d17095 100644
--- a/test/registered/vlm/test_patch_embed_perf.py
+++ b/test/registered/vlm/test_patch_embed_perf.py
@@ -8,7 +8,7 @@ import torch.nn as nn
 from sglang.srt.models.glm4v import Glm4vVisionPatchEmbed
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=12, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=8, suite="stage-b-test-1-gpu-large")
 
 PATCH_SIZE = 14
 TEMPORAL_PATCH_SIZE = 2
diff --git a/test/registered/vlm/test_video_utils.py b/test/registered/vlm/test_video_utils.py
index 5163f3519..f4dd0b027 100644
--- a/test/registered/vlm/test_video_utils.py
+++ b/test/registered/vlm/test_video_utils.py
@@ -5,7 +5,7 @@ import pytest
 from sglang.srt.utils import sample_video_frames
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+register_cpu_ci(est_time=4, suite="stage-a-test-cpu")
 
 
 class DummyVideo:
diff --git a/test/registered/vlm/test_vision_chunked_prefill.py b/test/registered/vlm/test_vision_chunked_prefill.py
index f078fab02..57739821d 100644
--- a/test/registered/vlm/test_vision_chunked_prefill.py
+++ b/test/registered/vlm/test_vision_chunked_prefill.py
@@ -1,6 +1,6 @@
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=150, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=143, suite="stage-b-test-1-gpu-large")
 register_amd_ci(est_time=270, suite="stage-b-test-1-gpu-small-amd")
 """
 Usage:
diff --git a/test/registered/vlm/test_vision_openai_server_a.py b/test/registered/vlm/test_vision_openai_server_a.py
index f64f7b8bc..c41eb1f61 100644
--- a/test/registered/vlm/test_vision_openai_server_a.py
+++ b/test/registered/vlm/test_vision_openai_server_a.py
@@ -19,7 +19,7 @@ from sglang.test.vlm_utils import (
     VideoOpenAITestMixin,
 )
 
-register_cuda_ci(est_time=700, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=736, suite="stage-b-test-1-gpu-large")
 
 
 class TestLlavaServer(ImageOpenAITestMixin):
diff --git a/test/registered/vlm/test_vlm_input_format.py b/test/registered/vlm/test_vlm_input_format.py
index 44af1e5d8..792da109e 100644
--- a/test/registered/vlm/test_vlm_input_format.py
+++ b/test/registered/vlm/test_vlm_input_format.py
@@ -37,7 +37,7 @@ from sglang.srt.entrypoints.openai.protocol import ChatCompletionRequest
 from sglang.srt.parser.conversation import generate_chat_conv
 from sglang.srt.utils.hf_transformers_utils import _fix_added_tokens_encoding
 
-register_cuda_ci(est_time=620, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=602, suite="stage-b-test-1-gpu-large")
 
 IMAGE_MAN_IRONING_URL = "https://raw.githubusercontent.com/sgl-project/sgl-test-files/refs/heads/main/images/man_ironing_on_back_of_suv.png"
 IMAGE_SGL_LOGO_URL = "https://raw.githubusercontent.com/sgl-project/sgl-test-files/refs/heads/main/images/sgl_logo.png"
diff --git a/test/registered/vlm/test_vlm_tp4.py b/test/registered/vlm/test_vlm_tp4.py
index 6f1a12394..33137dc00 100644
--- a/test/registered/vlm/test_vlm_tp4.py
+++ b/test/registered/vlm/test_vlm_tp4.py
@@ -15,7 +15,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=140, suite="stage-c-test-4-gpu-h100")
+register_cuda_ci(est_time=118, suite="stage-c-test-4-gpu-h100")
 
 QWEN35_27B_MODEL = "Qwen/Qwen3.5-27B"
 MMMU_ACCURACY_THRESHOLD = 0.65