ci: B200 conditional split + LPT_SLOP removal (stage-c partition 8→3) (#25203)

scripts/ci/utils/compute_partitions.py

@@ -39,15 +39,13 @@ _STAGE_A_OVERRIDES = {
     "stage-a-test-1-gpu-small": 1,
 }
 
-# Per-partition wall-clock target + ceiling. Single knob for the whole
-# pipeline. ~17 min avg under perfect LPT (TARGET / LPT_SLOP), ~22 min under
-# worst-case LPT 4/3 imbalance, fail-fast above 30 min.
+# Per-partition wall-clock target. ~20 min avg naive; worst-case LPT 4/3
+# imbalance is ~27 min, still below the 30-min job-level timeout that acts
+# as the real safety net. No LPT slop applied — we lean on the runtime
+# timeout + the explicit MAX_PARTITION_SECONDS sanity check rather than
+# padding partition count.
 TARGET_SECONDS = 20 * 60
 
-# LPT (Longest Processing Time first) worst case is 4/3 * OPT; pad ~15% so a
-# slightly-unlucky LPT result still fits inside MAX_PARTITION_SECONDS.
-LPT_SLOP = 1.15
-
 # Hard ceiling. Exceeded → raise, forcing the maintainer to split a slow file
 # or bump TARGET_SECONDS deliberately.
 MAX_PARTITION_SECONDS = 30 * 60
@@ -97,10 +95,11 @@ def compute_partitions(tests, full_parallel=False):
             size = _STAGE_A_OVERRIDES[suite]
             max_parallel = size
         else:
-            size = max(1, math.ceil(total * LPT_SLOP / TARGET_SECONDS))
+            size = max(1, math.ceil(total / TARGET_SECONDS))
             max_parallel = size if full_parallel else compute_max_parallel(size)
-        # Check naive average (total/size). LPT can be ~4/3 of that, but the
-        # ceil + LPT_SLOP padding above absorbs that slack.
+        # Check naive average (total/size). LPT can be ~4/3 of that in
+        # worst case; the 30-min job timeout enforces the real ceiling at
+        # runtime. This build-time check fails fast on egregious misconfigs.
         if total / size > MAX_PARTITION_SECONDS:
             raise RuntimeError(
                 f"Suite {suite!r}: total est_time {total:.0f}s / size {size} "

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=392, stage="stage-c", runner_config="4-gpu-h100")
-register_cuda_ci(est_time=740, stage="stage-c", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=350, stage="stage-c", runner_config="4-gpu-b200")
 
 
 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, stage="stage-c", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=540, 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, stage="stage-c", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=340, stage="stage-c", runner_config="4-gpu-b200")
 
 QWEN35_FP4_MODEL = "nvidia/Qwen3.5-397B-A17B-NVFP4"
 ACC_THRESHOLDS = {QWEN35_FP4_MODEL: {"gsm8k": 0.95}}

test/registered/4-gpu-models/test_qwen35_models.py

@@ -17,7 +17,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=260, stage="stage-c", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=370, stage="stage-c", runner_config="4-gpu-b200")
 
 QWEN35_FP4_MODEL = "nvidia/Qwen3.5-397B-A17B-NVFP4"
 ACC_THRESHOLDS = {QWEN35_FP4_MODEL: {"gsm8k": 0.95}}

test/registered/attention/test_flash_attention_4.py

@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
 )
 
 # FlashAttention4 integration test (requires SM 100+ / Blackwell B200)
-register_cuda_ci(est_time=265, stage="stage-b", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=260, stage="stage-b", runner_config="4-gpu-b200")
 
 
 @unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")

test/registered/lora/test_lora_gpt_oss_20b_logprob_diff.py

@@ -35,9 +35,9 @@ from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(
-    est_time=300,
-    stage="stage-c",
-    runner_config="4-gpu-b200",
+    est_time=90,
+    suite="nightly-4-gpu-b200",
+    nightly=True,
 )
 
 BASE_MODEL = "lmsys/gpt-oss-20b-bf16"

test/registered/lora/test_lora_nemotron_3_super_120b_a12b_logprob_diff.py

@@ -35,9 +35,9 @@ from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(
-    est_time=300,
-    stage="stage-c",
-    runner_config="4-gpu-b200",
+    est_time=100,
+    suite="nightly-4-gpu-b200",
+    nightly=True,
 )
 
 BASE_MODEL = "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16"

test/registered/lora/test_lora_qwen3_30b_a3b_instruct_2507_logprob_diff.py

@@ -35,9 +35,9 @@ from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(
-    est_time=160,
-    stage="stage-c",
-    runner_config="4-gpu-b200",
+    est_time=100,
+    suite="nightly-4-gpu-b200",
+    nightly=True,
 )
 
 BASE_MODEL = "Qwen/Qwen3-30B-A3B-Instruct-2507"

test/registered/lora/test_lora_qwen3_5_35b_a3b_logprob_diff.py

@@ -35,7 +35,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(
-    est_time=160,
+    est_time=110,
     stage="stage-c",
     runner_config="4-gpu-b200",
 )

test/registered/lora/test_lora_qwen3_vl_30b_a3b_instruct_logprob_diff.py

@@ -35,7 +35,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(
-    est_time=160,
+    est_time=110,
     stage="stage-c",
     runner_config="4-gpu-b200",
 )

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, stage="stage-c", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=24, suite="nightly-4-gpu-b200", nightly=True)
 
 SKIP_TEST = torch.cuda.get_device_capability() < (10, 0)
 SKIP_REASON = "Nvfp4 Requires compute capability of 10 or above."

test/registered/quant/test_deepseek_v32_fp4_mtp_4gpu.py

@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
 )
 
 register_cuda_ci(
-    est_time=1060,
+    est_time=690,
     stage="stage-c",
     runner_config="4-gpu-b200",
 )

test/registered/quant/test_deepseek_v3_fp4_4gpu.py

@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=1190, stage="stage-c", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=960, stage="stage-c", runner_config="4-gpu-b200")
 
 FULL_DEEPSEEK_V3_FP4_MODEL_PATH = "nvidia/DeepSeek-V3-0324-FP4"
 SERVER_LAUNCH_TIMEOUT = 1200

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, stage="stage-c", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=430, 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=420, stage="stage-c", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=350, stage="stage-c", runner_config="4-gpu-b200")
 
 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=400, stage="stage-c", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=320, suite="nightly-4-gpu-b200", nightly=True)
 
 import unittest
 

test/registered/spec/eagle/test_deepseek_v3_fp4_mtp_small.py

@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=420, stage="stage-b", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=340, stage="stage-b", runner_config="4-gpu-b200")
 
 FULL_DEEPSEEK_V3_FP4_MODEL_PATH = "nvidia/DeepSeek-V3-0324-FP4"
 SERVER_LAUNCH_TIMEOUT = 1200

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=123, stage="stage-c", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=90, stage="stage-c", runner_config="4-gpu-b200")
 
 
 def test_gsm8k(base_url: str, model: str):