[None][test] add Perf sanity gb200 test into QA test db (NVIDIA#11882)

Signed-off-by: Xin He (SW-GPU) <200704525+xinhe-nv@users.noreply.github.com>

Author: xinhe-nv

jenkins/scripts/perf/local/submit.py

@@ -321,7 +321,7 @@ def generate_srun_args(args, runtime_mode, timestamp):
 
 
 def generate_pytest_command(
-    llm_src, work_dir, config_file_base_name, select_pattern, runtime_mode, benchmark_mode
+    test_prefix, work_dir, config_file_base_name, select_pattern, runtime_mode, benchmark_mode
 ):
     """Generate pytest command and test list."""
     # Generate test list content based on runtime_mode and benchmark_mode
@@ -344,8 +344,8 @@ def generate_pytest_command(
     test_list_path = os.path.join(work_dir, "test_list.txt")
 
     pytest_command = (
-        f"pytest -v -s "
-        f"--test-prefix={llm_src}/tests/integration/defs "
+        f"pytest -v "
+        f"--test-prefix={test_prefix} "
         f"--test-list={test_list_path} "
         f"--output-dir={work_dir} "
         f"-o junit_logging=out-err"
@@ -423,6 +423,7 @@ def main():
         default="1-100",
         help="Nsys start-stop range for generation workers in disaggregated mode (default: 1-100)",
     )
+    parser.add_argument("--test-prefix", default="", help="Test prefix")
 
     args = parser.parse_args()
 
@@ -480,6 +481,8 @@ def main():
         work_dir = os.path.join(llm_src, "jenkins", "scripts", "perf", "local", timestamp)
     os.makedirs(work_dir, exist_ok=True)
 
+    test_prefix = args.test_prefix if args.test_prefix else f"{llm_src}/tests/integration/defs"
+
     # Determine paths
     launch_sh = args.launch_sh if args.launch_sh else os.path.join(work_dir, "slurm_launch.sh")
     run_sh = (
@@ -521,7 +524,7 @@ def main():
 
     # Generate pytest command
     pytest_command, test_list_content, test_list_path = generate_pytest_command(
-        llm_src, work_dir, config_file_base_name, select_pattern, runtime_mode, benchmark_mode
+        test_prefix, work_dir, config_file_base_name, select_pattern, runtime_mode, benchmark_mode
     )
 
     # Write test list file

tests/integration/test_lists/qa/llm_perf_multinode.yml

@@ -5,11 +5,30 @@ llm_perf_multinode:
 - condition:
     wildcards:
       gpu:
-      - '*b200*'
+      - 'b200'
   tests:
   - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-b200_deepseek-r1-fp4_1k1k_con1_ctx1_dep4_gen1_tep8_eplb0_mtp3_ccb-UCX] TIMEOUT (120)
   - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-b200_deepseek-r1-fp4_1k1k_con2048_ctx1_dep4_gen1_dep8_eplb0_mtp1_ccb-UCX] TIMEOUT (120)
   - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-b200_deepseek-r1-fp4_1k1k_con256_ctx1_dep4_gen1_dep8_eplb0_mtp3_ccb-UCX] TIMEOUT (120)
   - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-b200_deepseek-r1-fp4_8k1k_con1536_ctx1_dep4_gen1_dep8_eplb0_mtp1_ccb-UCX] TIMEOUT (120)
   - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-b200_deepseek-r1-fp4_8k1k_con1_ctx1_dep4_gen1_tep8_eplb0_mtp3_ccb-UCX] TIMEOUT (120)
   - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-b200_deepseek-r1-fp4_8k1k_con256_ctx1_dep4_gen1_dep8_eplb0_mtp1_ccb-UCX] TIMEOUT (120)
+
+# 2: GB200 test cases
+- condition:
+    wildcards:
+      gpu:
+      - 'gb200'
+  tests:
+  - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-gb200_gpt-oss-120b-fp4_1k1k_con2048_ctx1_tp1_gen1_dep2_eplb0_mtp0_ccb-UCX] TIMEOUT (120)
+  - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-gb200_gpt-oss-120b-fp4_1k1k_con512_ctx1_tp1_gen1_dep2_eplb0_mtp0_ccb-UCX] TIMEOUT (120)
+  - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-gb200_gpt-oss-120b-fp4_1k1k_con64_ctx1_tp1_gen1_tp4_eplb0_mtp0_ccb-UCX] TIMEOUT (120)
+  - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-gb200_gpt-oss-120b-fp4_8k1k_con128_ctx1_tp1_gen1_tp4_eplb0_mtp0_ccb-UCX] TIMEOUT (120)
+  - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-gb200_gpt-oss-120b-fp4_8k1k_con4_ctx1_tp1_gen1_tp4_eplb0_mtp0_ccb-UCX] TIMEOUT (120)
+  - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-gb200_deepseek-r1-fp4_1k1k_con3072_ctx1_dep4_gen1_dep4_eplb0_mtp1_ccb-UCX] TIMEOUT (120)
+  - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-gb200_deepseek-v32-fp4_1k1k_con2048_ctx1_dep4_gen1_dep4_eplb0_mtp1_ccb-UCX] TIMEOUT (120)
+  - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-gb200_kimi-k2-thinking-fp4_1k1k_con4_ctx1_dep4_gen1_tep4_eplb0_mtp0_ccb-UCX] TIMEOUT (120)
+  - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-gb200_deepseek-r1-fp4_1k1k_con1024_ctx1_dep4_gen1_dep8_eplb0_mtp0_ccb-UCX] TIMEOUT (120)
+  - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-gb200_deepseek-r1-fp4_1k1k_con1_ctx1_dep4_gen1_tep8_eplb0_mtp3_ccb-UCX] TIMEOUT (120)
+  - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-gb200_deepseek-r1-fp4_8k1k_con1_ctx1_dep4_gen1_tep8_eplb0_mtp3_ccb-UCX] TIMEOUT (120)
+  - perf/test_perf_sanity.py::test_e2e[disagg-gen_only-gb200_deepseek-r1-fp4_128k8k_con1_ctx1_pp8_gen1_tep8_eplb0_mtp3_ccb-UCX] TIMEOUT (120)