[Xeon] CPU CI enhancement for Intel Xeon platforms (#24649)

.github/workflows/pr-test-xeon.yml

@@ -29,10 +29,18 @@ jobs:
     runs-on: ubuntu-latest
     outputs:
       main_package: ${{ steps.filter.outputs.main_package || steps.run-mode.outputs.run_all_tests}}
+    env:
+      PYTHONDONTWRITEBYTECODE: 1
     steps:
+      - name: clean up
+        run: |
+          sudo rm -rf ${{ github.workspace }} || true
+          sudo mkdir -p ${{ github.workspace }}
+          sudo chown -R $(id -u):$(id -g) ${{ github.workspace }}
       - name: Checkout code
         uses: actions/checkout@v4
         with:
+          clean: false
           ref: ${{ inputs.ref || github.ref }}
 
       - name: Determine run mode
@@ -72,16 +80,27 @@ jobs:
   build-test:
     needs: [check-changes, pr-gate]
     if: needs.check-changes.outputs.main_package == 'true'
-    runs-on: xeon-gnr
-    env:
-      HF_HOME: /home/sdp/.cache/huggingface
+    runs-on: ${{ matrix.runner }}
+
     strategy:
       matrix:
-        build_type: ['all']
+        include:
+          - runner: xeon-gnr
+            suite: base-b-test-cpu
+            partition_args: "--auto-partition-id 0 --auto-partition-size 2"
+          - runner: xeon-gnr
+            suite: base-b-test-cpu
+            partition_args: "--auto-partition-id 1 --auto-partition-size 2"
     steps:
+      - name: Cleanup workspace
+        run: |
+          sudo rm -rf ${{ github.workspace }} || true
+          sudo mkdir -p ${{ github.workspace }}
+          sudo chown $(id -u):$(id -g) ${{ github.workspace }}
       - name: Checkout repository
         uses: actions/checkout@v4
         with:
+          clean: false
           ref: ${{ inputs.ref || github.ref }}
 
       - name: Build and Push
@@ -100,27 +119,28 @@ jobs:
         run: |
           docker run -dt \
             -v ${{ github.workspace }}:/sglang-checkout/ --ipc=host \
-            -v ${HF_HOME}:/root/.cache/huggingface \
-            --name ci_sglang_xeon \
+            -v $HOME/.cache/huggingface:/root/.cache/huggingface \
+            -e HF_TOKEN="$(cat ~/huggingface_token.txt)" \
+            --name ci_sglang_${{ matrix.runner }} \
             sglang_xeon
 
       - name: Check AMX support
         id: check_amx
         timeout-minutes: 5
         run: |
-          docker exec -w /sglang-checkout/ ci_sglang_xeon \
+          docker exec -w /sglang-checkout/ ci_sglang_${{ matrix.runner }} \
             bash -c "source /opt/.venv/bin/activate && python3 -c 'import torch; import sgl_kernel; assert torch._C._cpu._is_amx_tile_supported(); assert hasattr(torch.ops.sgl_kernel, \"convert_weight_packed\"); '"
 
       - name: Run unit tests
         timeout-minutes: 36
         run: |
-          docker exec -w /sglang-checkout/ ci_sglang_xeon \
-            bash -c "source /opt/.venv/bin/activate && cd ./test && python3 run_suite.py --hw cpu --suite base-b-test-cpu"
+          docker exec -w /sglang-checkout/ ci_sglang_${{ matrix.runner }} \
+            bash -c "source /opt/.venv/bin/activate && cd ./test && python3 run_suite.py --hw cpu --suite ${{ matrix.suite }} ${{ matrix.partition_args }}"
 
       - name: Change permission
         timeout-minutes: 2
         run: |
-          docker exec -u root ci_sglang_xeon bash -c "
+          docker exec -u root ci_sglang_${{ matrix.runner }} bash -c "
             rm -rf /tmp/ci-home  &&
             chown -R  $(id -u):$(id -g) /sglang-checkout/ 2>/dev/null || true
           "
@@ -128,4 +148,4 @@ jobs:
       - name: Cleanup container
         if: always()
         run: |
-          docker rm -f ci_sglang_xeon || true
+          docker rm -f ci_sglang_${{ matrix.runner }} || true

test/registered/bench_fn/test_benchmark_datasets_api.py

@@ -27,6 +27,7 @@ from sglang.benchmark.datasets.sharegpt import sample_sharegpt_requests
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=6, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class _DummyTokenTensor:

test/registered/debug_utils/comparator/aligner/entrypoint/test_executor.py

@@ -34,6 +34,7 @@ from sglang.srt.debug_utils.comparator.utils import Pair
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=15, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 class TestExecuteSubPlans:

test/registered/debug_utils/comparator/aligner/entrypoint/test_planner.py

@@ -26,6 +26,7 @@ from sglang.srt.debug_utils.comparator.utils import Pair
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=15, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 def _make_meta(

test/registered/debug_utils/comparator/aligner/reorderer/test_executor.py

@@ -27,6 +27,7 @@ from sglang.srt.debug_utils.comparator.dims_spec import (
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 def _zigzag_order(cp_size: int) -> list[int]:

test/registered/debug_utils/comparator/aligner/reorderer/test_planner.py

@@ -27,6 +27,7 @@ from sglang.srt.debug_utils.comparator.dims_spec import (
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 class TestComputeReordererPlans:

test/registered/debug_utils/comparator/aligner/test_axis_aligner.py

@@ -18,6 +18,7 @@ from sglang.srt.debug_utils.comparator.utils import Pair
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=15, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 class TestComputeAxisAlignerPlan:

test/registered/debug_utils/comparator/aligner/token_aligner/test_aux_plugins.py

@@ -17,6 +17,7 @@ from sglang.srt.debug_utils.comparator.dims_spec import TokenLayout
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=15, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 _sglang_plugin = _SGLangPlugin()
 _megatron_plugin = _MegatronPlugin()

test/registered/debug_utils/comparator/aligner/token_aligner/test_concat_steps.py

@@ -11,6 +11,7 @@ from sglang.srt.debug_utils.comparator.utils import Pair
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=15, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 class TestExecuteConcat:

test/registered/debug_utils/comparator/aligner/token_aligner/test_executor.py

@@ -30,6 +30,7 @@ from sglang.srt.debug_utils.comparator.utils import Pair
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=15, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 def _named(tensor: torch.Tensor, names: list[str]) -> torch.Tensor:

test/registered/debug_utils/comparator/aligner/token_aligner/test_planner.py

@@ -24,6 +24,7 @@ from sglang.srt.debug_utils.comparator.utils import Pair
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=30, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 class TestBuildTokenIndexSGLangThd:

test/registered/debug_utils/comparator/aligner/unsharder/test_executor.py

@@ -31,6 +31,7 @@ from sglang.srt.debug_utils.comparator.output_types import ReplicatedCheckResult
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 def _name_tensors(

test/registered/debug_utils/comparator/aligner/unsharder/test_parallel_info.py

@@ -10,6 +10,7 @@ from sglang.srt.debug_utils.comparator.dims_spec import ParallelAxis
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 class TestNormalizeParallelInfo:

test/registered/debug_utils/comparator/aligner/unsharder/test_planner.py

@@ -20,6 +20,7 @@ from sglang.srt.debug_utils.comparator.dims_spec import ParallelAxis, parse_dims
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 class TestComputeUnsharderPlan:

test/registered/debug_utils/comparator/dims_spec/test_dim_parser.py

@@ -13,6 +13,7 @@ from sglang.srt.debug_utils.comparator.dims_spec import (
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=5, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 class TestParseDim:

test/registered/debug_utils/comparator/dims_spec/test_dims_parser.py

@@ -16,6 +16,7 @@ from sglang.srt.debug_utils.comparator.dims_spec import (
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=5, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 class TestSingletonDimUtilFilterOut:

test/registered/debug_utils/comparator/dims_spec/test_tensor_naming.py

@@ -15,6 +15,7 @@ from sglang.srt.debug_utils.comparator.dims_spec import (
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=5, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 class TestFindDimIndex:

test/registered/debug_utils/comparator/dims_spec/test_types.py

@@ -10,6 +10,7 @@ from sglang.srt.debug_utils.comparator.dims_spec import (
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=5, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 class TestDimConstants:

test/registered/debug_utils/comparator/tensor_comparator/test_comparator.py

@@ -15,6 +15,7 @@ from sglang.srt.debug_utils.comparator.tensor_comparator.types import DiffInfo
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=20, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 class TestComputeTensorInfo:

test/registered/debug_utils/comparator/tensor_comparator/test_types.py

@@ -22,6 +22,7 @@ from sglang.srt.debug_utils.comparator.tensor_comparator.types import (
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 def _make_stats(**overrides) -> TensorStats:

test/registered/debug_utils/comparator/test_log_sink.py

@@ -12,6 +12,7 @@ from sglang.srt.debug_utils.comparator.report_sink import report_sink
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 def _make_error_log(**overrides) -> ErrorLog:

test/registered/debug_utils/comparator/test_meta_overrider.py

@@ -17,6 +17,7 @@ from sglang.srt.debug_utils.comparator.meta_overrider import (
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 # ───────────────────── Unit: MetaOverrideRule ─────────────────────

test/registered/debug_utils/comparator/test_model_validation.py

@@ -45,6 +45,7 @@ from sglang.srt.debug_utils.comparator.utils import Pair, _check_equal_lengths
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 class TestCheckEqualLengths:

test/registered/debug_utils/comparator/test_preset.py

@@ -4,6 +4,7 @@ from sglang.srt.debug_utils.comparator.preset import PRESETS, expand_preset
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=5, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 class TestExpandPreset:

test/registered/debug_utils/comparator/test_utils.py

@@ -18,6 +18,7 @@ from sglang.srt.debug_utils.comparator.utils import (
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=1, suite="base-b-test-cpu")
 
 
 class TestCalcRelDiff:

test/registered/debug_utils/source_patcher/test_code_patcher.py

@@ -11,6 +11,7 @@ from sglang.srt.debug_utils.source_patcher.types import EditSpec, PatchSpec
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 SAMPLE_MODULE_NAME = "_source_patcher_test_fixtures.sample_module"
 

test/registered/debug_utils/source_patcher/test_dumper_integration.py

@@ -9,6 +9,7 @@ from sglang.srt.debug_utils.dumper import DumperConfig, _Dumper
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=8, suite="base-b-test-cpu")
 
 SAMPLE_MODULE_NAME = "_source_patcher_test_fixtures.sample_module"
 

test/registered/debug_utils/source_patcher/test_source_editor.py

@@ -6,6 +6,7 @@ from sglang.srt.debug_utils.source_patcher.types import EditSpec, PatchApplicati
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class TestApplyEdits:

test/registered/debug_utils/test_crash_dump.py

@@ -9,7 +9,11 @@ import requests
 
 from sglang.srt.environ import envs
 from sglang.srt.utils import kill_process_tree
-from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.ci.ci_register import (
+    register_amd_ci,
+    register_cpu_ci,
+    register_cuda_ci,
+)
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
@@ -19,6 +23,7 @@ from sglang.test.test_utils import (
 
 register_cuda_ci(est_time=40, suite="nightly-1-gpu", nightly=True)
 register_amd_ci(est_time=40, suite="nightly-amd-1-gpu", nightly=True)
+register_cpu_ci(est_time=225, suite="base-b-test-cpu")
 
 
 class TestCrashDump(CustomTestCase):

test/registered/debug_utils/test_schedule_simulator.py

@@ -26,7 +26,7 @@ from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cpu_ci(est_time=120, suite="base-a-test-cpu", nightly=True)
-
+register_cpu_ci(est_time=338, suite="base-b-test-cpu")
 
 # ==================== Non-E2E Tests ====================
 

test/registered/debug_utils/test_soft_watchdog.py

@@ -5,7 +5,11 @@ import requests
 
 from sglang.srt.environ import envs
 from sglang.srt.utils import kill_process_tree
-from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.ci.ci_register import (
+    register_amd_ci,
+    register_cpu_ci,
+    register_cuda_ci,
+)
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
@@ -15,6 +19,7 @@ from sglang.test.test_utils import (
 
 register_cuda_ci(est_time=120, suite="nightly-1-gpu", nightly=True)
 register_amd_ci(est_time=120, suite="nightly-amd-1-gpu", nightly=True)
+register_cpu_ci(est_time=622, suite="base-b-test-cpu")
 
 
 class BaseTestSoftWatchdog:

test/registered/function_call/test_kimik2_detector.py

@@ -12,6 +12,7 @@ from sglang.srt.parser.reasoning_parser import KimiK2Detector as KimiK2Reasoning
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(5, "base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 def _make_tool(name, parameters=None):

test/registered/lora/test_lora_eviction_policy.py

@@ -20,10 +20,15 @@ Tests LRU and FIFO eviction behavior.
 import unittest
 
 from sglang.srt.lora.eviction_policy import get_eviction_policy
-from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.ci.ci_register import (
+    register_amd_ci,
+    register_cpu_ci,
+    register_cuda_ci,
+)
 
 register_cuda_ci(est_time=200, suite="nightly-1-gpu", nightly=True)
 register_amd_ci(est_time=200, suite="nightly-amd-1-gpu", nightly=True)
+register_cpu_ci(est_time=6, suite="base-b-test-cpu")
 
 
 class TestLoRAEvictionPolicy(unittest.TestCase):

test/registered/lora/test_lora_openai_api.py

@@ -10,10 +10,15 @@ from unittest.mock import MagicMock
 
 from sglang.srt.entrypoints.openai.serving_base import OpenAIServingBase
 from sglang.srt.server_args import ServerArgs
-from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.ci.ci_register import (
+    register_amd_ci,
+    register_cpu_ci,
+    register_cuda_ci,
+)
 
 register_cuda_ci(est_time=30, suite="nightly-1-gpu", nightly=True)
 register_amd_ci(est_time=30, suite="nightly-amd-1-gpu", nightly=True)
+register_cpu_ci(est_time=8, suite="base-b-test-cpu")
 
 
 class MockTokenizerManager:

test/registered/model_loading/test_external_models.py

@@ -2,11 +2,16 @@ import unittest
 
 import sglang as sgl
 from sglang.srt.environ import envs
-from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.ci.ci_register import (
+    register_amd_ci,
+    register_cpu_ci,
+    register_cuda_ci,
+)
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=29, stage="base-b", runner_config="1-gpu-small")
 register_amd_ci(est_time=45, suite="stage-b-test-1-gpu-small-amd")
+register_cpu_ci(est_time=203, suite="base-b-test-cpu")
 
 
 class TestExternalModels(CustomTestCase):

test/registered/models/test_transformers_backend_eval.py

@@ -3,11 +3,12 @@
 import unittest
 from types import SimpleNamespace
 
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_cpu_ci, 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=48, stage="base-b", runner_config="1-gpu-small")
+register_cpu_ci(est_time=320, suite="base-b-test-cpu")
 
 
 class TestTransformersBackendEval(DefaultServerBase):

test/registered/observability/test_priority_metrics.py

@@ -8,7 +8,11 @@ from prometheus_client.samples import Sample
 
 from sglang.srt.observability.metrics_collector import QueueCount
 from sglang.srt.utils import kill_process_tree
-from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.ci.ci_register import (
+    register_amd_ci,
+    register_cpu_ci,
+    register_cuda_ci,
+)
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
@@ -22,6 +26,7 @@ register_cuda_ci(
     runner_config="1-gpu-small",
 )
 register_amd_ci(est_time=60, suite="stage-b-test-1-gpu-small-amd")
+register_cpu_ci(est_time=179, suite="base-b-test-cpu")
 
 _MODEL_NAME = "Qwen/Qwen3-0.6B"
 

test/registered/quant/test_quant_config_parsing.py

@@ -6,6 +6,7 @@ from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cpu_ci(est_time=15, suite="base-a-test-cpu")
+register_cpu_ci(est_time=8, suite="base-b-test-cpu")
 
 
 class TestQuantLogString(CustomTestCase):

test/registered/rl/test_pause_generation_tensor_consistency.py

@@ -10,6 +10,7 @@ from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cpu_ci(est_time=7, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 # ---------------------------------------------------------------------------

test/registered/scheduler/test_abort_with_metrics.py

@@ -19,6 +19,7 @@ from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cpu_ci(est_time=7, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 _HTTP_SCOPE = {
     "type": "http",

test/registered/scheduler/test_routing_key_scheduling.py

@@ -6,7 +6,11 @@ import unittest
 import aiohttp
 
 from sglang.srt.utils import kill_process_tree
-from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.ci.ci_register import (
+    register_amd_ci,
+    register_cpu_ci,
+    register_cuda_ci,
+)
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
@@ -18,6 +22,7 @@ from sglang.test.test_utils import (
 
 register_cuda_ci(est_time=120, suite="nightly-1-gpu", nightly=True)
 register_amd_ci(est_time=120, suite="nightly-amd-1-gpu", nightly=True)
+register_cpu_ci(est_time=184, suite="base-b-test-cpu")
 
 
 class TestRoutingKeyScheduling(CustomTestCase):

test/registered/unit/constrained/test_grammar_manager.py

@@ -28,6 +28,7 @@ from sglang.srt.constrained.reasoner_grammar_backend import ReasonerGrammarObjec
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(2.0, "base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 def _make_scheduler(grammar_backend_name="none", skip_tokenizer=False):

test/registered/unit/constrained/test_reasoner_grammar_backend.py

@@ -16,6 +16,7 @@ from sglang.srt.constrained.torch_ops.token_filter_torch_ops import (
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(2.0, "base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class _DummyTokenizer:

test/registered/unit/constrained/test_utils.py

@@ -15,6 +15,7 @@ from sglang.srt.constrained.utils import is_legacy_structural_tag
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(1.0, "base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class TestIsLegacyStructuralTag(unittest.TestCase):

test/registered/unit/entrypoints/openai/test_serving_embedding.py

@@ -57,6 +57,7 @@ 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="base-a-test-cpu")
+register_cpu_ci(est_time=8, suite="base-b-test-cpu")
 
 
 # Mock TokenizerManager for embedding tests

test/registered/unit/function_call/test_function_call_parser.py

@@ -32,6 +32,7 @@ from sglang.srt.function_call.qwen3_coder_detector import Qwen3CoderDetector
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=15, suite="base-a-test-cpu")
+register_cpu_ci(est_time=61, suite="base-b-test-cpu")
 
 
 class TestPythonicDetector(unittest.TestCase):

test/registered/unit/function_call/test_json_schema_constraint.py

@@ -19,6 +19,7 @@ from sglang.srt.function_call.utils import (
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(5, "base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class TestJsonSchemaConstraint(unittest.TestCase):

test/registered/unit/function_call/test_parallel_tool_calls.py

@@ -24,6 +24,7 @@ from sglang.srt.function_call.json_array_parser import JsonArrayParser
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(5, "base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class TestParallelToolCalls(unittest.TestCase):

test/registered/unit/function_call/test_unknown_tool_name.py

@@ -10,6 +10,7 @@ from sglang.srt.function_call.core_types import StreamingParseResult
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(5, "base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class DummyDetector(BaseFormatDetector):

test/registered/unit/managers/test_io_struct.py

@@ -2,7 +2,11 @@ import copy
 import unittest
 
 from sglang.srt.managers.io_struct import GenerateReqInput
-from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.ci.ci_register import (
+    register_amd_ci,
+    register_cpu_ci,
+    register_cuda_ci,
+)
 from sglang.test.test_utils import (
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
     DEFAULT_URL_FOR_TEST,
@@ -11,6 +15,7 @@ from sglang.test.test_utils import (
 
 register_cuda_ci(est_time=8, stage="base-b", runner_config="1-gpu-large")
 register_amd_ci(est_time=8, suite="stage-b-test-1-gpu-small-amd")
+register_cpu_ci(est_time=8, suite="base-b-test-cpu")
 
 
 class TestGenerateReqInputNormalization(CustomTestCase):

test/registered/unit/managers/test_prefill_adder.py

@@ -9,11 +9,16 @@ from sglang.srt.mem_cache.base_prefix_cache import (
     IncLockRefResult,
 )
 from sglang.srt.server_args import ServerArgs, set_global_server_args_for_scheduler
-from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.ci.ci_register import (
+    register_amd_ci,
+    register_cpu_ci,
+    register_cuda_ci,
+)
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=9, stage="base-b", runner_config="1-gpu-small")
 register_amd_ci(est_time=2, suite="stage-b-test-1-gpu-small-amd")
+register_cpu_ci(est_time=8, suite="base-b-test-cpu")
 
 
 class TestPrefillAdder(CustomTestCase):

test/registered/unit/managers/test_profile_merger_http_api.py

@@ -2,10 +2,15 @@ import json
 import unittest
 
 from sglang.srt.managers.io_struct import ProfileReqInput
-from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.ci.ci_register import (
+    register_amd_ci,
+    register_cpu_ci,
+    register_cuda_ci,
+)
 
 register_cuda_ci(est_time=8, stage="base-b", runner_config="1-gpu-small")
 register_amd_ci(est_time=9, suite="stage-b-test-1-gpu-small-amd")
+register_cpu_ci(est_time=8, suite="base-b-test-cpu")
 
 
 class TestProfileMergerHTTPAPI(unittest.TestCase):

test/registered/unit/managers/test_scheduler_flush_cache.py

@@ -13,6 +13,7 @@ from sglang.srt.managers.scheduler_components.flush_wrapper import (
 )
 
 register_cpu_ci(est_time=14, suite="base-a-test-cpu")
+register_cpu_ci(est_time=8, suite="base-b-test-cpu")
 
 
 class TestSchedulerFlushCache(unittest.TestCase):

test/registered/unit/managers/test_scheduler_pause_generation.py

@@ -12,6 +12,7 @@ from sglang.srt.managers.scheduler import Scheduler
 from sglang.srt.managers.scheduler_components.pool_stats_observer import PoolStats
 
 register_cpu_ci(est_time=15, suite="base-a-test-cpu")
+register_cpu_ci(est_time=9, suite="base-b-test-cpu")
 
 
 class TestSchedulerPauseGeneration(unittest.TestCase):

test/registered/unit/mem_cache/test_evict_policy.py

@@ -3,6 +3,7 @@
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=6, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 import unittest
 from unittest.mock import MagicMock

test/registered/unit/mem_cache/test_radix_cache_slru_accuracy.py

@@ -12,10 +12,15 @@ from sglang.srt.mem_cache.base_prefix_cache import (
 from sglang.srt.mem_cache.cache_init_params import CacheInitParams
 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_amd_ci, register_cuda_ci
+from sglang.test.ci.ci_register import (
+    register_amd_ci,
+    register_cpu_ci,
+    register_cuda_ci,
+)
 
 register_cuda_ci(est_time=8, stage="base-b", runner_config="1-gpu-small")
 register_amd_ci(est_time=8, suite="stage-b-test-1-gpu-small-amd")
+register_cpu_ci(est_time=8, suite="base-b-test-cpu")
 
 
 class TestSLRUAccuracy(unittest.TestCase):

test/registered/unit/models/test_llava.py

@@ -2,11 +2,16 @@ import unittest
 from unittest.mock import patch
 
 from sglang.srt.models.llava import AutoModel, LlavaForConditionalGeneration
-from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.ci.ci_register import (
+    register_amd_ci,
+    register_cpu_ci,
+    register_cuda_ci,
+)
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=9, stage="base-b", runner_config="1-gpu-small")
 register_amd_ci(est_time=9, suite="stage-b-test-1-gpu-small-amd")
+register_cpu_ci(est_time=8, suite="base-b-test-cpu")
 
 
 class PixtralVisionConfig:

test/registered/unit/observability/test_cpu_monitor.py

@@ -7,6 +7,7 @@ from unittest.mock import MagicMock, patch
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=60, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class TestCpuMonitor(unittest.TestCase):

test/registered/unit/observability/test_func_timer.py

@@ -3,6 +3,7 @@
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=6, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 import asyncio
 import unittest

test/registered/unit/observability/test_label_transform.py

@@ -3,6 +3,7 @@
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=6, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 import unittest
 

test/registered/unit/observability/test_metrics_utils.py

@@ -7,6 +7,7 @@ from sglang.srt.observability.utils import (
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=6, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class TestMetricsUtils(unittest.TestCase):

test/registered/unit/observability/test_request_metrics_exporter.py

@@ -6,6 +6,7 @@ from typing import Any, Dict, List, Optional
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=6, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 import asyncio
 import json

test/registered/unit/observability/test_startup_func_log_and_timer.py

@@ -3,6 +3,7 @@
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=6, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 import unittest
 from unittest.mock import MagicMock, patch

test/registered/unit/parser/test_code_completion_parser.py

@@ -19,6 +19,7 @@ from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cpu_ci(est_time=7, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class TestFimPosition(CustomTestCase):

test/registered/unit/parser/test_conversation.py

@@ -33,6 +33,7 @@ from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cpu_ci(est_time=7, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class TestConversationGetPrompt(CustomTestCase):

test/registered/unit/parser/test_harmony_parser.py

@@ -15,6 +15,7 @@ from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cpu_ci(est_time=7, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class TestEvent(CustomTestCase):

test/registered/unit/parser/test_jinja_template_utils.py

@@ -10,6 +10,7 @@ from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cpu_ci(est_time=7, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class TestTemplateContentFormatDetection(CustomTestCase):

test/registered/unit/parser/test_reasoning_content_without_parser.py

@@ -5,6 +5,7 @@ from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cpu_ci(est_time=7, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 # Simulated model output that contains think tags (e.g. from DeepSeek-R1)
 THINK_OUTPUT = (

test/registered/unit/sampling/test_custom_logit_processor.py

@@ -3,6 +3,7 @@
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=7, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 import json
 import unittest

test/registered/unit/sampling/test_penaltylib.py

@@ -3,6 +3,7 @@
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=9, suite="base-a-test-cpu")
+register_cpu_ci(est_time=8, suite="base-b-test-cpu")
 
 import unittest
 from unittest.mock import MagicMock

test/registered/unit/sampling/test_sampling_batch_info.py

@@ -3,6 +3,7 @@
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=9, suite="base-a-test-cpu")
+register_cpu_ci(est_time=8, suite="base-b-test-cpu")
 
 import unittest
 from unittest.mock import MagicMock, patch

test/registered/unit/sampling/test_sampling_params.py

@@ -3,6 +3,7 @@
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=7, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 import unittest
 from unittest.mock import MagicMock

test/registered/unit/server_args/test_server_args.py

@@ -12,6 +12,7 @@ from sglang.test.test_utils import (
 )
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu")
+register_cpu_ci(est_time=12, suite="base-b-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")

test/registered/unit/test_runai_utils.py

@@ -7,6 +7,7 @@ from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cpu_ci(est_time=7, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class TestRunaiUtils(CustomTestCase):

test/registered/unit/utils/test_gauge_histogram.py

@@ -1,6 +1,7 @@
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=6, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 import unittest
 

test/registered/unit/utils/test_http_server_auth.py

@@ -11,6 +11,7 @@ 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=6, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class TestHttpServerAdminAuth(unittest.TestCase):

test/registered/unit/utils/test_json_response.py

@@ -11,6 +11,7 @@ from sglang.srt.utils.json_response import (
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=6, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class TestJSONResponseUtils(unittest.TestCase):

test/registered/unit/utils/test_patch_tokenizer.py

@@ -12,6 +12,7 @@ from sglang.srt.utils.patch_tokenizer import (
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=30, suite="base-a-test-cpu", nightly=True)
+register_cpu_ci(est_time=53, suite="base-b-test-cpu")
 
 
 class TestPatchTokenizerEndToEndTest(unittest.TestCase):

test/registered/unit/utils/test_profile_merger.py

@@ -15,10 +15,15 @@ import unittest
 
 from sglang.srt.managers.io_struct import ProfileReq, ProfileReqInput, ProfileReqType
 from sglang.srt.utils.profile_merger import ProfileMerger
-from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.ci.ci_register import (
+    register_amd_ci,
+    register_cpu_ci,
+    register_cuda_ci,
+)
 
 register_cuda_ci(est_time=9, stage="base-b", runner_config="1-gpu-small")
 register_amd_ci(est_time=8, suite="stage-b-test-1-gpu-small-amd")
+register_cpu_ci(est_time=8, suite="base-b-test-cpu")
 
 
 class TestProfileMerger(unittest.TestCase):

test/registered/unit/utils/test_subprocess_watchdog.py

@@ -25,6 +25,7 @@ from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cpu_ci(est_time=9, suite="base-a-test-cpu")
+register_cpu_ci(est_time=9, suite="base-b-test-cpu")
 
 
 def healthy_worker():

test/registered/utils/test_log_utils.py

@@ -11,6 +11,7 @@ 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=6, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 _LOG_PREFIX_RE = re.compile(r"^\[\d{4}-\d{2}-\d{2} \d{2}:\d{2}:\d{2}\] ")
 

test/registered/utils/test_model_file_verifier.py

@@ -20,7 +20,7 @@ from sglang.srt.utils.model_file_verifier import (
     generate_checksums,
     verify,
 )
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_cpu_ci, register_cuda_ci
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
@@ -29,6 +29,7 @@ from sglang.test.test_utils import (
 
 # Note: AMD registration removed - test_model_file_verifier fails on AMD
 register_cuda_ci(est_time=120, suite="nightly-1-gpu", nightly=True)
+register_cpu_ci(est_time=540, suite="base-b-test-cpu")
 
 MODEL_NAME = "Qwen/Qwen3-0.6B"
 

test/registered/utils/test_network_address.py

@@ -7,6 +7,7 @@ from sglang.srt.utils.network import NetworkAddress
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=7, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-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")

test/registered/utils/test_socket_utils.py

@@ -16,6 +16,7 @@ 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="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class TestTryBindSocket(CustomTestCase):

test/registered/utils/test_type_based_dispatcher.py

@@ -8,10 +8,11 @@ import timeit
 import unittest
 
 from sglang.srt.managers.io_struct import SamplingParams
-from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cpu_ci
 from sglang.utils import TypeBasedDispatcher
 
 register_amd_ci(est_time=10, suite="stage-b-test-1-gpu-small-amd")
+register_cpu_ci(est_time=8, suite="base-b-test-cpu")
 
 
 class TestTypeBasedDispatcher(unittest.TestCase):

test/registered/vlm/test_evs.py

@@ -3,11 +3,16 @@ from types import SimpleNamespace
 
 import pytest
 
-from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.ci.ci_register import (
+    register_amd_ci,
+    register_cpu_ci,
+    register_cuda_ci,
+)
 from sglang.test.test_utils import run_doctests
 
 register_cuda_ci(est_time=11, stage="base-b", runner_config="1-gpu-small")
 register_amd_ci(est_time=20, suite="stage-b-test-1-gpu-small-amd")
+register_cpu_ci(est_time=8, suite="base-b-test-cpu")
 
 
 def test_resolve_evs_config():

test/registered/vlm/test_video_utils.py

@@ -6,6 +6,7 @@ from sglang.srt.utils import sample_video_frames
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=7, suite="base-a-test-cpu")
+register_cpu_ci(est_time=7, suite="base-b-test-cpu")
 
 
 class DummyVideo:

test/srt/cpu/test_cpu_graph.py

@@ -28,7 +28,7 @@ class TestCPUGraph(CustomTestCase):
             "--batch-size",
             "1",
             "--mem-fraction-static",
-            "0.05",
+            "0.5",
             "--enable-torch-compile",
             "--torch-compile-max-bs",
             "2",
@@ -55,7 +55,7 @@ class TestCPUGraph(CustomTestCase):
                 "--attention-backend",
                 "intel_amx",
                 "--mem-fraction-static",
-                "0.05",
+                "0.5",
                 "--disable-radix",
                 "--trust-remote-code",
                 "--disable-overlap-schedule",