[ci] split stage-c-test-4-gpu-b200 to enable a low-disk runner pool (#23417)

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

test/registered/4-gpu-models/test_gpt_oss_4gpu.py

@@ -4,7 +4,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.gpt_oss_common import BaseTestGptOss
 
 register_cuda_ci(est_time=328, suite="stage-c-test-4-gpu-h100")
-register_cuda_ci(est_time=740, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=740, suite="stage-c-test-4-gpu-b200-small")
 
 
 class TestGptOss4Gpu(BaseTestGptOss):

test/registered/4-gpu-models/test_nvidia_nemotron_3_super_nvfp4.py

@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=710, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=710, suite="nightly-4-gpu-b200", nightly=True)
 
 NEMOTRON_3_SUPER_NVFP4_MODEL = "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4"
 

test/registered/4-gpu-models/test_qwen35_fp4_mtp_v2.py

@@ -15,7 +15,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=540, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=540, suite="stage-c-test-4-gpu-b200-small")
 
 QWEN35_FP4_MODEL = "nvidia/Qwen3.5-397B-A17B-NVFP4"
 ACC_THRESHOLDS = {QWEN35_FP4_MODEL: {"gsm8k": 0.95}}

test/registered/4-gpu-models/test_qwen35_fp4_triton.py

@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
     ModelLaunchSettings,
 )
 
-register_cuda_ci(est_time=720, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=720, suite="stage-c-test-4-gpu-b200-small")
 
 QWEN35_FP4_MODEL = "nvidia/Qwen3.5-397B-A17B-NVFP4"
 ACC_THRESHOLDS = {QWEN35_FP4_MODEL: {"gsm8k": 0.95}}

test/registered/lora/test_lora_gpt_oss_20b_logprob_diff.py

@@ -36,7 +36,7 @@ from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(
     est_time=300,
-    suite="stage-c-test-4-gpu-b200",
+    suite="stage-c-test-4-gpu-b200-small",
 )
 
 BASE_MODEL = "lmsys/gpt-oss-20b-bf16"

test/registered/lora/test_lora_qwen3_30b_a3b_instruct_2507_logprob_diff.py

@@ -36,7 +36,7 @@ from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(
     est_time=160,
-    suite="stage-c-test-4-gpu-b200",
+    suite="stage-c-test-4-gpu-b200-small",
 )
 
 BASE_MODEL = "Qwen/Qwen3-30B-A3B-Instruct-2507"

test/registered/lora/test_lora_qwen3_vl_30b_a3b_instruct_logprob_diff.py

@@ -36,7 +36,7 @@ from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(
     est_time=160,
-    suite="stage-c-test-4-gpu-b200",
+    suite="stage-c-test-4-gpu-b200-small",
 )
 
 BASE_MODEL = "Qwen/Qwen3-VL-30B-A3B-Instruct"

test/registered/moe/test_cutedsl_moe.py

@@ -16,7 +16,7 @@ except ImportError:
     CuteDslMoEWrapper = None
     convert_sf_to_mma_layout = None
 
-register_cuda_ci(est_time=590, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=590, suite="stage-c-test-4-gpu-b200-small")
 
 SKIP_TEST = torch.cuda.get_device_capability() < (10, 0)
 SKIP_REASON = "Nvfp4 Requires compute capability of 10 or above."

test/registered/quant/test_fp8_blockwise_gemm.py

@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=630, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=630, suite="nightly-4-gpu-b200", nightly=True)
 
 MODEL_PATH = "Qwen/Qwen3-4B-Instruct-2507-FP8"
 MXFP8_MODEL_PATH = "zianglih/Qwen3-4B-Instruct-2507-MXFP8"

test/registered/quant/test_nvfp4_gemm.py

@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=550, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=550, suite="nightly-4-gpu-b200", nightly=True)
 
 MODEL_PATH = "nvidia/Llama-3.1-8B-Instruct-NVFP4"
 

test/registered/rl/test_update_weights_from_disk_blackwell.py

@@ -1,6 +1,6 @@
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=149, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=149, suite="stage-c-test-4-gpu-b200-small")
 
 import unittest
 

test/registered/spec/eagle/test_eagle_infer_beta_dp_attention.py

@@ -17,7 +17,7 @@ from sglang.test.test_utils import (
 )
 
 # EAGLE with DP attention on B200 (tp=2, dp=2, requires 4 B200 GPUs)
-register_cuda_ci(est_time=136, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=136, suite="stage-c-test-4-gpu-b200-small")
 
 
 def test_gsm8k(base_url: str, model: str):

test/run_suite.py

@@ -48,6 +48,7 @@ PER_COMMIT_SUITES = {
         "stage-b-kernel-benchmark-1-gpu-large",
         "stage-c-test-4-gpu-h100",
         "stage-c-test-4-gpu-b200",
+        "stage-c-test-4-gpu-b200-small",
         "stage-c-test-4-gpu-gb200",
         "stage-c-test-8-gpu-h20",
         "stage-c-test-8-gpu-h200",