From 46c2b7762765f512ee7697794d83b9304d47d01c Mon Sep 17 00:00:00 2001
From: Kangyan-Zhou <zky314343421@gmail.com>
Date: Wed, 8 Apr 2026 17:04:57 -0700
Subject: [PATCH] [CI] Add GLM-5.1 nightly tests and update Qwen3.5 model
 (#22399)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 .../8-gpu-models/test_glm_51_fp8.py           | 69 +++++++++++++++++++
 test/registered/8-gpu-models/test_qwen35.py   | 13 +++-
 test/registered/gb300/test_glm5_fp8.py        |  6 +-
 3 files changed, 82 insertions(+), 6 deletions(-)
 create mode 100644 test/registered/8-gpu-models/test_glm_51_fp8.py

diff --git a/test/registered/8-gpu-models/test_glm_51_fp8.py b/test/registered/8-gpu-models/test_glm_51_fp8.py
new file mode 100644
index 000000000..ace31a06f
--- /dev/null
+++ b/test/registered/8-gpu-models/test_glm_51_fp8.py
@@ -0,0 +1,69 @@
+import unittest
+
+from sglang.test.accuracy_test_runner import AccuracyTestParams
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.performance_test_runner import PerformanceTestParams
+from sglang.test.run_combined_tests import run_combined_tests
+from sglang.test.test_utils import ModelLaunchSettings
+
+# Runs on both H200 and B200 via nightly-8-gpu-common suite
+register_cuda_ci(est_time=1800, suite="nightly-8-gpu-common", nightly=True)
+
+GLM_51_FP8_MODEL_PATH = "zai-org/GLM-5.1-FP8"
+
+COMMON_ARGS = [
+    "--trust-remote-code",
+    "--reasoning-parser=glm45",
+    "--tool-call-parser=glm47",
+    "--mem-fraction-static=0.9",
+    "--enable-metrics",
+]
+
+MTP_ARGS = [
+    "--speculative-algorithm=EAGLE",
+    "--speculative-num-steps=3",
+    "--speculative-eagle-topk=1",
+    "--speculative-num-draft-tokens=4",
+]
+
+
+class TestGlm51Fp8(unittest.TestCase):
+    """GLM-5.1 FP8 on H200/B200 (8-GPU, tp=8)."""
+
+    def test_glm51_fp8(self):
+        dp_args = ["--dp=8", "--enable-dp-attention"]
+
+        variants = [
+            ModelLaunchSettings(
+                GLM_51_FP8_MODEL_PATH,
+                tp_size=8,
+                extra_args=COMMON_ARGS,
+                variant="TP8",
+            ),
+            ModelLaunchSettings(
+                GLM_51_FP8_MODEL_PATH,
+                tp_size=8,
+                extra_args=COMMON_ARGS + dp_args,
+                variant="TP8+DP8",
+            ),
+            ModelLaunchSettings(
+                GLM_51_FP8_MODEL_PATH,
+                tp_size=8,
+                extra_args=COMMON_ARGS + dp_args + MTP_ARGS,
+                variant="TP8+DP8+MTP",
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
+            ),
+        ]
+
+        run_combined_tests(
+            models=variants,
+            test_name="GLM-5.1-FP8",
+            accuracy_params=AccuracyTestParams(dataset="gsm8k", baseline_accuracy=0.92),
+            performance_params=PerformanceTestParams(
+                profile_dir="performance_profiles_glm_51_fp8",
+            ),
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/test/registered/8-gpu-models/test_qwen35.py b/test/registered/8-gpu-models/test_qwen35.py
index 7386b4a8c..813552b83 100644
--- a/test/registered/8-gpu-models/test_qwen35.py
+++ b/test/registered/8-gpu-models/test_qwen35.py
@@ -9,7 +9,7 @@ from sglang.test.test_utils import ModelLaunchSettings
 # Runs on both H200 and B200 via nightly-8-gpu-common suite
 register_cuda_ci(est_time=1800, suite="nightly-8-gpu-common", nightly=True)
 
-QWEN35_MODEL_PATH = "Qwen/Qwen3.5-397B-A17B"
+QWEN35_MODEL_PATH = "Qwen/Qwen3.5-397B-A17B-FP8"
 
 
 class TestQwen35(unittest.TestCase):
@@ -30,6 +30,7 @@ class TestQwen35(unittest.TestCase):
             "--tool-call-parser=qwen3_coder",
             "--mem-fraction-static=0.8",
         ]
+        dp_args = ["--dp=8", "--enable-dp-attention"]
         mtp_args = [
             "--speculative-algorithm=EAGLE",
             "--speculative-num-steps=3",
@@ -48,8 +49,14 @@ class TestQwen35(unittest.TestCase):
             ModelLaunchSettings(
                 QWEN35_MODEL_PATH,
                 tp_size=8,
-                extra_args=base_args + mtp_args,
-                variant="TP8+MTP",
+                extra_args=base_args + dp_args,
+                variant="TP8+DP8",
+            ),
+            ModelLaunchSettings(
+                QWEN35_MODEL_PATH,
+                tp_size=8,
+                extra_args=base_args + dp_args + mtp_args,
+                variant="TP8+DP8+MTP",
                 env={"SGLANG_ENABLE_SPEC_V2": "1"},
             ),
         ]
diff --git a/test/registered/gb300/test_glm5_fp8.py b/test/registered/gb300/test_glm5_fp8.py
index e429e5873..388f21b63 100644
--- a/test/registered/gb300/test_glm5_fp8.py
+++ b/test/registered/gb300/test_glm5_fp8.py
@@ -8,7 +8,7 @@ from sglang.test.test_utils import ModelLaunchSettings
 
 register_cuda_ci(est_time=7200, suite="nightly-4-gpu-gb300", nightly=True)
 
-MODEL_PATH = "zai-org/GLM-5-FP8"
+MODEL_PATH = "zai-org/GLM-5.1-FP8"
 
 COMMON_ARGS = [
     "--trust-remote-code",
@@ -27,7 +27,7 @@ MTP_ARGS = [
 
 
 class TestGlm5Fp8(unittest.TestCase):
-    """GLM-5 FP8 on GB300 (4x B200 NVL4, tp=4)."""
+    """GLM-5.1 FP8 on GB300 (4x B200 NVL4, tp=4)."""
 
     def test_glm5_fp8(self):
         variants = [
@@ -56,7 +56,7 @@ class TestGlm5Fp8(unittest.TestCase):
 
         run_combined_tests(
             models=variants,
-            test_name="GLM-5-FP8",
+            test_name="GLM-5.1-FP8",
             accuracy_params=AccuracyTestParams(dataset="gsm8k", baseline_accuracy=0.92),
             performance_params=PerformanceTestParams(
                 profile_dir="performance_profiles_gb300",