[NPU CI] Reorganize test output/log directory structure with workflow context (#33685)

2026-08-18 23:46:00 +08:00
parent 63d783bbe0
commit 499e90a125
139 changed files with 779 additions and 688 deletions
@@ -7,10 +7,9 @@ from sglang.test.ascend.test_ascend_utils import DEEPSEEK_V3_2_EXP_W8A8_WEIGHTS_
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
-    suite="",
+    est_time=4800,
+    suite="nightly-acc-16-npu-a3",
    nightly=True,
-    disabled="accuracy testcase",
 )

 OTHER_ARGS = [
@@ -7,7 +7,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import GLM_4_6V_FLASH_MOD
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
+    est_time=4800,
    suite="",
    nightly=True,
    disabled="performance testcase",
@@ -7,10 +7,9 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import GLM_4_7_FLASH_MODE
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
-    suite="",
+    est_time=6500,
+    suite="nightly-acc-2-npu-a3",
    nightly=True,
-    disabled="accuracy testcase",
 )

 ENVS = {
@@ -7,6 +7,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import GLM_4_7_FLASH_MODE
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(est_time=3600, suite="base-c-test-acc-2-npu-a3")
+register_npu_ci(est_time=6500, suite="nightly-acc-2-npu-a3", nightly=True)

 ENVS = {
    "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
@@ -10,7 +10,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
+    est_time=4800,
    suite="",
    nightly=True,
    disabled="accuracy testcase",
@@ -9,6 +9,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(est_time=3600, suite="base-c-test-acc-16-npu-a3")
+register_npu_ci(est_time=4800, suite="nightly-acc-16-npu-a3", nightly=True)

 ENVS = {
    "SGLANG_SET_CPU_AFFINITY": "1",
@@ -12,7 +12,7 @@ from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
    est_time=1800,
-    suite="nightly-8-npu-a3",
+    suite="full-8-npu-a3",
    nightly=True,
    disabled="accuracy testcase",
 )
@@ -11,10 +11,9 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
-    suite="",
+    est_time=7200,
+    suite="nightly-acc-16-npu-a3",
    nightly=True,
-    disabled="accuracy testcase",
 )

 MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_ENVS = {
@@ -11,10 +11,9 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
-    suite="full-16-npu-a3",
+    est_time=4800,
+    suite="nightly-acc-16-npu-a3",
    nightly=True,
-    disabled="accuracy testcase",
 )

 MINIMAX_M2_5_HIGH_THROUGHPUT_ENVS = {
@@ -9,6 +9,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(est_time=3600, suite="base-c-test-acc-2-npu-a3")
+register_npu_ci(est_time=1300, suite="nightly-acc-2-npu-a3", nightly=True)

 MODEL_ENVS = {
    "SGLANG_SET_CPU_AFFINITY": "1",
@@ -10,10 +10,9 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
-    suite="",
+    est_time=3700,
+    suite="nightly-acc-2-npu-a3",
    nightly=True,
-    disabled="accuracy testcase",
 )

 QWEN3_8B_ENVS = {
@@ -10,10 +10,9 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
-    suite="",
+    est_time=3700,
+    suite="nightly-acc-2-npu-a3",
    nightly=True,
-    disabled="accuracy testcase",
 )

 QWEN3_8B_ENVS = {
@@ -10,10 +10,9 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
-    suite="",
+    est_time=2800,
+    suite="nightly-acc-2-npu-a3",
    nightly=True,
-    disabled="accuracy testcase",
 )

 QWEN3_30B_A3B_ENVS = {
@@ -10,10 +10,9 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
-    suite="",
+    est_time=4800,
+    suite="nightly-acc-16-npu-a3",
    nightly=True,
-    disabled="performance testcase",
 )

 QWEN3_32B_ENVS = {
@@ -10,10 +10,9 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
-    suite="",
+    est_time=4800,
+    suite="nightly-acc-4-npu-a3",
    nightly=True,
-    disabled="accuracy testcase",
 )

 QWEN3_32B_ENVS = {
@@ -10,10 +10,9 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
-    suite="",
+    est_time=4800,
+    suite="nightly-1-npu-a2",
    nightly=True,
-    disabled="accuracy testcase",
 )

 QWEN3_32B_ENVS = {
@@ -9,6 +9,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(est_time=3600, suite="base-c-test-acc-2-npu-a3")
+register_npu_ci(est_time=2800, suite="nightly-acc-2-npu-a3", nightly=True)

 QWEN3_5_9B_ENVS = {
    "SGLANG_SET_CPU_AFFINITY": "1",
@@ -9,10 +9,13 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
-    suite="",
+    # Measured ~45min/round in CI (2026-08-17 nightly-acc-2-npu-a3).
+    # est_time = 2x single-round so derive_timeout_per_file (1.5x est)
+    # reserves ~3 rounds of budget for accuracy retries.
+    # Normalized to the shared 6500 ladder value for ~42-54min cases.
+    est_time=6500,
+    suite="nightly-acc-2-npu-a3",
    nightly=True,
-    disabled="performance testcase",
 )

 QWEN3_6_27B_64K_PREFIX_ENVS = {
@@ -9,10 +9,9 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
-    suite="full-2-npu-a3",
+    est_time=8400,
+    suite="nightly-acc-2-npu-a3",
    nightly=True,
-    disabled="accuracy testcase",
 )

 QWEN3_6_27B_3K5_1K5_ENVS = {
@@ -8,12 +8,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 )
 from sglang.test.ci.ci_register import register_npu_ci

-register_npu_ci(
-    est_time=3600,
-    suite="full-2-npu-a3",
-    nightly=True,
-    disabled="performance testcase",
-)
+register_npu_ci(est_time=6500, suite="nightly-acc-2-npu-a3", nightly=True)

 QWEN3_6_35B_A3B_3K5_1K5_ENVS = {
    "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
@@ -9,10 +9,9 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
-    suite="",
+    est_time=2800,
+    suite="nightly-acc-2-npu-a3",
    nightly=True,
-    disabled="accuracy testcase",
 )

 QWEN3_6_35B_A3B_64K_PREFIX_ENVS = {
@@ -10,10 +10,9 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
-    suite="",
+    est_time=4800,
+    suite="nightly-acc-4-npu-a3",
    nightly=True,
-    disabled="accuracy testcase",
 )

 QWEN3_NEXT_80B_A3B_ENVS = {
@@ -9,7 +9,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
+    est_time=4800,
    suite="",
    nightly=True,
    disabled="performance testcase",
@@ -9,6 +9,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(est_time=3600, suite="base-c-test-acc-4-npu-a3")
+register_npu_ci(est_time=4800, suite="nightly-acc-4-npu-a3", nightly=True)

 QWEN3_VL_30B_A3B_ENVS = {
    "SGLANG_SET_CPU_AFFINITY": "1",
@@ -9,10 +9,9 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
-    est_time=3600,
-    suite="",
+    est_time=8400,
+    suite="nightly-acc-2-npu-a3",
    nightly=True,
-    disabled="performance testcase",
 )

 ENVS = {
@@ -9,6 +9,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(est_time=3600, suite="base-c-test-acc-4-npu-a3")
+register_npu_ci(est_time=4800, suite="nightly-acc-4-npu-a3", nightly=True)

 QWEN3_VL_8B_ENVS = {
    "SGLANG_SET_CPU_AFFINITY": "1",
@@ -10,6 +10,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(est_time=3600, suite="base-c-test-perf-2-npu-a3")
+register_npu_ci(est_time=6500, suite="nightly-acc-2-npu-a3", nightly=True)

 _is_pr_pipeline = os.environ.get("GITHUB_EVENT_NAME") == "pull_request"

@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
    popen_launch_server,
 )

-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestNPUHierarchicalCache(CustomTestCase):
@@ -10,7 +10,7 @@ from sglang.test.test_utils import CustomTestCase

 register_npu_ci(
    est_time=400,
-    suite="nightly-16-npu-a3",
+    suite="full-16-npu-a3",
    nightly=True,
 )

@@ -12,7 +12,7 @@ from sglang.test.test_utils import (

 register_npu_ci(
    est_time=400,
-    suite="nightly-1-npu-a3",
+    suite="full-1-npu-a3",
    nightly=True,
 )

@@ -9,7 +9,7 @@ from sglang.test.test_utils import CustomTestCase

 register_npu_ci(
    est_time=400,
-    suite="nightly-2-npu-a3",
+    suite="full-2-npu-a3",
    nightly=True,
 )

@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
    popen_launch_server,
 )

-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestNPURadixCache(CustomTestCase):
@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
 )

 register_npu_ci(est_time=400, suite="base-b-test-4-npu-a3")
-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)


 class TestLLaDA2Mini(GSM8KAscendMixin, CustomTestCase):
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
    run_bench_serving,
 )

-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestNoChunkedPrefill(CustomTestCase):
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
    popen_launch_server,
 )

-register_npu_ci(est_time=800, suite="nightly-2-npu-a3", nightly=True)
+register_npu_ci(est_time=800, suite="full-2-npu-a3", nightly=True)

 TEST_MODEL_MATRIX = {
    DEEPSEEK_CODER_V2_LITE_WEIGHTS_PATH,
@@ -7,7 +7,7 @@ from sglang.test.ascend.test_mmlu import TestMMLU
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-16-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-16-npu-a3", nightly=True)


 class TestDeepEpDeepseekV32(GSM8KAscendMixin, TestMMLU, CustomTestCase):
@@ -9,7 +9,7 @@ from sglang.test.ascend.test_mmlu import TestMMLU
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=200, suite="nightly-16-npu-a3", nightly=True)
+register_npu_ci(est_time=200, suite="full-16-npu-a3", nightly=True)


 class TestDeepEpQwen(GSM8KAscendMixin, TestMMLU, CustomTestCase):
@@ -11,7 +11,7 @@ from sglang.test.test_utils import CustomTestCase

 register_npu_ci(
    est_time=200,
-    suite="nightly-8-npu-a3",
+    suite="full-8-npu-a3",
    nightly=True,
 )

@@ -7,7 +7,7 @@ from sglang.test.ascend.test_mmlu import TestMMLU
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=200, suite="nightly-16-npu-a3", nightly=True)
+register_npu_ci(est_time=200, suite="full-16-npu-a3", nightly=True)


 class TestDeepEpDeepseekV32(GSM8KAscendMixin, TestMMLU, CustomTestCase):
@@ -9,7 +9,7 @@ from sglang.test.ascend.test_mmlu import TestMMLU
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=200, suite="nightly-16-npu-a3", nightly=True)
+register_npu_ci(est_time=200, suite="full-16-npu-a3", nightly=True)


 class TestDeepEpQwen(GSM8KAscendMixin, TestMMLU, CustomTestCase):
@@ -11,7 +11,7 @@ from sglang.test.test_utils import CustomTestCase

 register_npu_ci(
    est_time=200,
-    suite="nightly-8-npu-a3",
+    suite="full-8-npu-a3",
    nightly=True,
 )

@@ -18,7 +18,7 @@ from sglang.test.test_utils import (

 register_npu_ci(
    est_time=400,
-    suite="nightly-1-npu-a3",
+    suite="full-1-npu-a3",
    nightly=True,
 )

@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
    popen_launch_server,
 )

-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestLogLevel(CustomTestCase):
@@ -5,7 +5,7 @@ from sglang.test.test_utils import CustomTestCase, run_mmlu_test

 register_npu_ci(
    est_time=400,
-    suite="nightly-1-npu-a3",
+    suite="full-1-npu-a3",
    nightly=True,
 )

@@ -46,7 +46,7 @@ if torch.cuda.is_available():
    torch.backends.cuda.matmul.allow_tf32 = False
    torch.backends.cudnn.allow_tf32 = False

-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestOriginalLogprob(unittest.TestCase):
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (

 register_npu_ci(
    est_time=400,
-    suite="nightly-4-npu-a3",
+    suite="full-4-npu-a3",
    nightly=True,
 )

@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
    popen_launch_server,
 )

-register_npu_ci(est_time=300, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=300, suite="full-1-npu-a3", nightly=True)

 logger = logging.getLogger(__name__)

@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
    popen_launch_server,
 )

-register_npu_ci(est_time=500, suite="nightly-2-npu-a3", nightly=True)
+register_npu_ci(est_time=500, suite="full-2-npu-a3", nightly=True)

 logger = logging.getLogger(__name__)

@@ -22,7 +22,7 @@ def send_requests(url, **kwargs):
    responses.append(response)


-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestNpuApi(CustomTestCase):
@@ -20,7 +20,7 @@ logging.basicConfig(
 )
 logger = logging.getLogger(__name__)

-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestNpuApi(CustomTestCase):
@@ -15,7 +15,7 @@ from sglang.test.test_utils import (

 register_npu_ci(
    est_time=400,
-    suite="nightly-2-npu-a3",
+    suite="full-2-npu-a3",
    nightly=True,
 )

@@ -19,7 +19,7 @@ Each section should be as comprehensive as possible to create a rich and immersi
 The story should span multiple events, challenges, and character developments over time. Aim to make the story at least 3,000 words long.
 """

-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestMatchedStop(CustomTestCase):
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (

 register_npu_ci(
    est_time=400,
-    suite="nightly-1-npu-a3",
+    suite="full-1-npu-a3",
    nightly=True,
 )

@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
    popen_launch_server,
 )

-register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-2-npu-a3", nightly=True)


 class TestOpenAIServerIgnoreEOS(CustomTestCase):
@@ -15,7 +15,7 @@ from sglang.test.test_utils import (
    popen_launch_server,
 )

-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestPenalty(CustomTestCase):
@@ -4,7 +4,7 @@ from sglang.test.ascend.gsm8k_ascend_mixin import GSM8KAscendMixin
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-8-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-8-npu-a3", nightly=True)


 class TestDbrx(GSM8KAscendMixin, CustomTestCase):
@@ -7,7 +7,7 @@ from sglang.test.test_utils import CustomTestCase

 register_npu_ci(
    est_time=400,
-    suite="nightly-8-npu-a3",
+    suite="full-8-npu-a3",
    nightly=True,
 )

@@ -7,7 +7,7 @@ from sglang.test.test_utils import CustomTestCase

 register_npu_ci(
    est_time=400,
-    suite="nightly-2-npu-a3",
+    suite="full-2-npu-a3",
    nightly=True,
 )

@@ -5,7 +5,7 @@ from sglang.test.ascend.test_ascend_utils import BAICHUAN2_13B_CHAT_WEIGHTS_PATH
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestBaichuan(GSM8KAscendMixin, CustomTestCase):
@@ -5,7 +5,7 @@ from sglang.test.ascend.test_ascend_utils import CHATGLM2_6B_WEIGHTS_PATH
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestChatGlm2(GSM8KAscendMixin, CustomTestCase):
@@ -5,7 +5,7 @@ from sglang.test.ascend.test_ascend_utils import DEEPSEEK_V3_2_EXP_W8A8_WEIGHTS_
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-16-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-16-npu-a3", nightly=True)


 class TestDeepSeekV32(GSM8KAscendMixin, CustomTestCase):
@@ -7,7 +7,7 @@ from sglang.test.test_utils import CustomTestCase

 register_npu_ci(
    est_time=400,
-    suite="nightly-1-npu-a3",
+    suite="full-1-npu-a3",
    nightly=True,
 )

@@ -7,7 +7,7 @@ from sglang.test.test_utils import CustomTestCase

 register_npu_ci(
    est_time=400,
-    suite="nightly-1-npu-a3",
+    suite="full-1-npu-a3",
    nightly=True,
 )

@@ -8,7 +8,7 @@ from sglang.test.test_utils import CustomTestCase

 register_npu_ci(
    est_time=400,
-    suite="nightly-1-npu-a3",
+    suite="full-1-npu-a3",
    nightly=True,
 )

@@ -7,7 +7,7 @@ from sglang.test.ascend.test_ascend_utils import (
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-4-npu-a3", nightly=True)


 class TestLlama4(GSM8KAscendMixin, CustomTestCase):
@@ -5,7 +5,7 @@ from sglang.test.ascend.test_ascend_utils import LLAMA_2_7B_WEIGHTS_PATH
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestLlama(GSM8KAscendMixin, CustomTestCase):
@@ -5,7 +5,7 @@ from sglang.test.ascend.test_ascend_utils import LLAMA_2_7B_WEIGHTS_PATH
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-2-npu-a3", nightly=True)


 class TestLlama(GSM8KAscendMixin, CustomTestCase):
@@ -5,7 +5,7 @@ from sglang.test.ascend.test_ascend_utils import MINICPM3_4B_WEIGHTS_PATH
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestMiniCPM3(GSM8KAscendMixin, CustomTestCase):
@@ -5,7 +5,7 @@ from sglang.test.ascend.test_ascend_utils import MISTRAL_7B_INSTRUCT_V0_2_WEIGHT
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestMistral7B(GSM8KAscendMixin, CustomTestCase):
@@ -5,7 +5,7 @@ from sglang.test.ascend.test_ascend_utils import PHI_4_MULTIMODAL_INSTRUCT_WEIGH
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestPhi4(GSM8KAscendMixin, CustomTestCase):
@@ -5,7 +5,7 @@ from sglang.test.ascend.test_ascend_utils import QWEN3_0_6B_WEIGHTS_PATH
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)


 class TestQwen306B(GSM8KAscendMixin, CustomTestCase):
@@ -5,7 +5,7 @@ from sglang.test.ascend.test_ascend_utils import QWEN3_235B_A22B_W8A8_WEIGHTS_PA
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-8-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-8-npu-a3", nightly=True)


 class TestQwen3235BA22BW8A8(GSM8KAscendMixin, CustomTestCase):
@@ -7,7 +7,7 @@ from sglang.test.ascend.test_ascend_utils import (
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-2-npu-a3", nightly=True)


 class TestQwen330B(GSM8KAscendMixin, CustomTestCase):
@@ -6,7 +6,7 @@ from sglang.test.ascend.test_ascend_utils import QWEN3_30B_A3B_WEIGHTS_PATH
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=500, suite="nightly-4-npu-a3", nightly=True)
+register_npu_ci(est_time=500, suite="full-4-npu-a3", nightly=True)


 class TestQwen330BAttnCP(GSM8KAscendMixin, CustomTestCase):
@@ -7,7 +7,7 @@ from sglang.test.test_utils import CustomTestCase

 register_npu_ci(
    est_time=400,
-    suite="nightly-4-npu-a3",
+    suite="full-4-npu-a3",
    nightly=True,
 )

@@ -5,7 +5,7 @@ from sglang.test.ascend.test_ascend_utils import QWEN3_8B_WEIGHTS_PATH
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-2-npu-a3", nightly=True)


 class TestQwen38BCommQuantization(GSM8KAscendMixin, CustomTestCase):
@@ -9,7 +9,7 @@ from sglang.test.test_utils import CustomTestCase

 register_npu_ci(
    est_time=400,
-    suite="nightly-16-npu-a3",
+    suite="full-16-npu-a3",
    nightly=True,
 )

@@ -6,7 +6,7 @@ from sglang.test.ascend.test_ascend_utils import QWQ_32B_W8A8_WEIGHTS_PATH
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase

-register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-2-npu-a3", nightly=True)


 class TestQWQ32BW8A8(GSM8KAscendMixin, CustomTestCase):
@@ -8,7 +8,8 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 )
 from sglang.test.ci.ci_register import register_npu_ci

-register_npu_ci(est_time=3600, suite="base-c-test-perf-16-npu-a3")
+register_npu_ci(est_time=1200, suite="base-c-test-perf-16-npu-a3")
+register_npu_ci(est_time=1200, suite="nightly-perf-16-npu-a3", nightly=True)

 # Environment variables for DSV4-Flash single-node PD-mix deployment.
 DEEPSEEK_V4_FLASH_W8A8_8P_ENVS = {
@@ -55,7 +56,7 @@ DEEPSEEK_V4_FLASH_W8A8_8P_OTHER_ARGS = [
    "--watchdog-timeout",
    9000,
    "--mem-fraction-static",
-    0.7,
+    0.6,
    "--prefill-max-requests",
    2,
    "--disable-radix-cache",
@@ -70,7 +70,7 @@ GLM_5_1_PD_SEP_PREFILL_ARGS = [
    "--served-model-name",
    "glm-5",
    "--chunked-prefill-size",
-    16384,
+    32768,
    "--max-prefill-tokens",
    180000,
    "--moe-a2a-backend",
@@ -151,6 +151,14 @@ GLM_5_1_PD_SEP_DECODE_ARGS = [
    "glm45",
    "--tool-call-parser",
    "glm47",
+    "--speculative-algorithm",
+    "NEXTN",
+    "--speculative-num-steps",
+    3,
+    "--speculative-eagle-topk",
+    1,
+    "--speculative-num-draft-tokens",
+    4,
 ]

 GLM_5_1_PD_SEP_MODEL_CONFIG = {
@@ -171,8 +179,8 @@ class TestNPUGLM5_1_W4A8_PD_SEP_In3k5_Out1k5(TestNpuPerfMultiNodePdSepTestCaseBa
    benchmark_tool = BENCHMARK_TOOL_DEFAULT
    dataset_type = AISBENCHMARK_DATASET_DEFAULT
    dataset_name = "random"
-    max_concurrency = 1
-    num_prompts = 1
+    max_concurrency = 32
+    num_prompts = 64
    input_len = 65536
    output_len = 1024
    random_range_ratio = 1
@@ -12,7 +12,7 @@ from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
    est_time=1800,
-    suite="nightly-8-npu-a3",
+    suite="full-8-npu-a3",
    nightly=True,
    disabled="Currently it is executed by the npu performance workflow.",
 )
@@ -10,7 +10,8 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 )
 from sglang.test.ci.ci_register import register_npu_ci

-register_npu_ci(est_time=1800, suite="base-c-test-perf-16-npu-a3")
+register_npu_ci(est_time=1200, suite="base-c-test-perf-16-npu-a3")
+register_npu_ci(est_time=1200, suite="nightly-perf-16-npu-a3", nightly=True)

 KIMI_K2_6_ENVS = {
    "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
@@ -10,6 +10,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(est_time=3600, suite="base-c-test-perf-8-npu-a3")
+register_npu_ci(est_time=3600, suite="nightly-perf-8-npu-a3", nightly=True)

 MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_ENVS = {
    "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
@@ -12,9 +12,8 @@ from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
    est_time=3600,
-    suite="full-16-npu-a3",
+    suite="nightly-perf-16-npu-a3",
    nightly=True,
-    disabled="performance testcase",
 )

 MINIMAX_M2_5_HIGH_THROUGHPUT_ENVS = {
@@ -10,6 +10,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(est_time=3600, suite="base-c-test-perf-2-npu-a3")
+register_npu_ci(est_time=3600, suite="nightly-perf-2-npu-a3", nightly=True)

 QWEN3_8B_ENVS = {
    "SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT": "600",
@@ -9,9 +9,8 @@ from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
    est_time=3600,
-    suite="",
+    suite="nightly-perf-2-npu-a3",
    nightly=True,
-    disabled="performance testcase",
 )

 QWEN3_8B_ENVS = {
@@ -10,6 +10,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(est_time=3600, suite="base-c-test-perf-16-npu-a3")
+register_npu_ci(est_time=3600, suite="nightly-perf-16-npu-a3", nightly=True)

 QWEN3_235B_ENVS = {
    "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
@@ -10,6 +10,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(est_time=3600, suite="base-c-test-perf-2-npu-a3")
+register_npu_ci(est_time=3600, suite="nightly-perf-2-npu-a3", nightly=True)

 QWEN3_30B_A3B_ENVS = {
    "ASCEND_LAUNCH_BLOCKING": "0",
@@ -11,9 +11,8 @@ from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
    est_time=3600,
-    suite="",
+    suite="nightly-perf-16-npu-a3",
    nightly=True,
-    disabled="performance testcase",
 )

 QWEN3_32B_ENVS = {
@@ -10,6 +10,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(est_time=3600, suite="base-c-test-perf-4-npu-a3")
+register_npu_ci(est_time=3600, suite="nightly-perf-4-npu-a3", nightly=True)

 QWEN3_32B_ENVS = {
    "SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT": "600",
@@ -11,9 +11,8 @@ from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
    est_time=3600,
-    suite="",
+    suite="nightly-1-npu-a2",
    nightly=True,
-    disabled="performance testcase",
 )

 QWEN3_32B_ENVS = {
@@ -9,6 +9,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(est_time=3600, suite="base-c-test-perf-16-npu-a3")
+register_npu_ci(est_time=3600, suite="nightly-perf-16-npu-a3", nightly=True)

 QWEN3_5_397B_A17B_ENVS = {
    "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
@@ -9,6 +9,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(est_time=3600, suite="base-c-test-perf-2-npu-a3")
+register_npu_ci(est_time=3600, suite="nightly-perf-2-npu-a3", nightly=True)

 QWEN3_6_27B_1024_ENVS = {
    "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
@@ -10,9 +10,8 @@ from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
    est_time=3600,
-    suite="full-2-npu-a3",
+    suite="nightly-perf-2-npu-a3",
    nightly=True,
-    disabled="performance testcase",
 )

 QWEN3_6_27B_1080P_ENVS = {
@@ -9,9 +9,8 @@ from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
    est_time=3600,
-    suite="",
+    suite="nightly-perf-4-npu-a3",
    nightly=True,
-    disabled="performance testcase",
 )

 QWEN3_6_27B_64K_PREFIX_ENVS = {
@@ -10,9 +10,8 @@ from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
    est_time=3600,
-    suite="full-2-npu-a3",
+    suite="nightly-perf-2-npu-a3",
    nightly=True,
-    disabled="performance testcase",
 )

 QWEN3_6_27B_3K5_1K5_ENVS = {
@@ -9,6 +9,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(est_time=3600, suite="base-c-test-perf-2-npu-a3")
+register_npu_ci(est_time=3600, suite="nightly-perf-4-npu-a3", nightly=True)

 QWEN3_6_27B_64K_1K_ENVS = {
    "STREAMS_PER_DEVICE": "32",
@@ -10,9 +10,8 @@ from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
    est_time=3600,
-    suite="",
+    suite="nightly-perf-4-npu-a3",
    nightly=True,
-    disabled="performance testcase",
 )

 QWEN3_6_27B_128K_ENVS = {
@@ -10,9 +10,8 @@ from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
    est_time=3600,
-    suite="",
+    suite="nightly-perf-4-npu-a3",
    nightly=True,
-    disabled="performance testcase",
 )

 QWEN3_6_27B_16K_1k_ENVS = {
@@ -10,9 +10,8 @@ from sglang.test.ci.ci_register import register_npu_ci

 register_npu_ci(
    est_time=3600,
-    suite="",
+    suite="nightly-perf-2-npu-a3",
    nightly=True,
-    disabled="performance testcase",
 )

 QWEN3_6_35B_A3B_128K_1K_ENVS = {
--- a/Show More
+++ b/Show More