chore: update CI test est_time values (#22565)

Co-authored-by: Cheng Wan <chwan@rice.edu> Co-authored-by: sglang-bot <sglang-bot@users.noreply.github.com> Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com>
2026-04-10 18:15:02 -07:00
parent b5e4ae7b1a
commit 265696b176
250 changed files with 251 additions and 251 deletions
@@ -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):
@@ -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"

@@ -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}}
@@ -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}}
@@ -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"

@@ -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"

@@ -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"

@@ -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"

@@ -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"

@@ -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"

@@ -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"
@@ -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"
@@ -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):
@@ -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"

@@ -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"

@@ -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")
@@ -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")


@@ -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

@@ -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")
@@ -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

@@ -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")
@@ -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")
@@ -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(
@@ -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")


@@ -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")


@@ -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")


@@ -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")


@@ -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")


@@ -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:
@@ -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")


@@ -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")


@@ -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")


@@ -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")
@@ -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()
@@ -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")


@@ -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")


@@ -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")


@@ -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")


@@ -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"


@@ -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):
@@ -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(
@@ -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")


@@ -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):
@@ -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):
@@ -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.")
@@ -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(
@@ -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(
@@ -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")


@@ -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")
@@ -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")


@@ -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
@@ -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"

@@ -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):
@@ -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()

@@ -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")


@@ -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
@@ -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")


@@ -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")


@@ -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):
@@ -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):
@@ -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.
@@ -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")


@@ -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

@@ -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")


@@ -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
@@ -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

@@ -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
@@ -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
@@ -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):
@@ -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",
@@ -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 = [
@@ -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")


@@ -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")


@@ -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):
@@ -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")


@@ -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",
@@ -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):
@@ -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")


@@ -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):
@@ -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")


@@ -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):
@@ -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")


@@ -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
@@ -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):
@@ -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",
@@ -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):
@@ -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"

@@ -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")


@@ -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):
@@ -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")


@@ -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()
@@ -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."
@@ -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()
@@ -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):
@@ -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):
@@ -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")


@@ -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):
@@ -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"
@@ -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")


 # ============================================================================
@@ -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):
--- a/Show More
+++ b/Show More