[AMD] CI - migrate perf test and fix stage-b-test-1-gpu-amd (#17340)

Co-authored-by: Bingxu Chen <bingxche@amd.com>
Co-authored-by: bingxche <Bingxu.Chen@amd.com>
Co-authored-by: michaelzhang-ai <michaelzhang.ai@users.noreply.github.com>

test/registered/perf/test_bench_one_batch_1gpu.py

@@ -1,6 +1,6 @@
 import unittest
 
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
@@ -12,6 +12,7 @@ from sglang.test.test_utils import (
 )
 
 register_cuda_ci(est_time=120, suite="stage-b-test-large-1-gpu-performance")
+register_amd_ci(est_time=120, suite="stage-b-test-large-1-gpu-performance-amd")
 
 
 class TestBenchOneBatch1GPU(CustomTestCase):

test/registered/perf/test_bench_one_batch_2gpu.py

@@ -1,6 +1,6 @@
 import unittest
 
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
     DEFAULT_MOE_MODEL_NAME_FOR_TEST,
@@ -12,6 +12,7 @@ from sglang.test.test_utils import (
 )
 
 register_cuda_ci(est_time=180, suite="stage-b-test-large-2-gpu-performance")
+register_amd_ci(est_time=630, suite="stage-b-test-large-2-gpu-performance-amd")
 
 
 class TestBenchOneBatch2GPU(CustomTestCase):

test/registered/perf/test_bench_serving_1gpu_large.py

@@ -4,7 +4,8 @@ Performance tests for single GPU that need H200 (80GB) - FP8 and EAGLE tests.
 
 import unittest
 
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.srt.utils import is_hip
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import (
     DEFAULT_DRAFT_MODEL_EAGLE,
     DEFAULT_MODEL_NAME_FOR_TEST_FP8,
@@ -17,6 +18,7 @@ from sglang.test.test_utils import (
 )
 
 register_cuda_ci(est_time=300, suite="stage-b-test-large-1-gpu-performance")
+register_amd_ci(est_time=300, suite="stage-b-test-large-1-gpu-performance-amd")
 
 
 class TestBenchServing1GPULarge(CustomTestCase):
@@ -38,6 +40,7 @@ class TestBenchServing1GPULarge(CustomTestCase):
             else:
                 self.assertGreater(res["output_throughput"], 4300)
 
+    @unittest.skipIf(is_hip(), "Skip Eagle test for ROCm")
     def test_online_latency_eagle(self):
         res = run_bench_serving(
             model=DEFAULT_TARGET_MODEL_EAGLE,

test/registered/perf/test_bench_serving_1gpu_part1.py

@@ -9,7 +9,7 @@ import unittest
 
 import requests
 
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
     CustomTestCase,
@@ -20,6 +20,7 @@ from sglang.test.test_utils import (
 )
 
 register_cuda_ci(est_time=1000, suite="stage-b-test-large-1-gpu-performance")
+register_amd_ci(est_time=1100, suite="stage-b-test-large-1-gpu-performance-amd")
 
 
 class TestBenchServing1GPUPart1(CustomTestCase):

test/registered/perf/test_bench_serving_1gpu_part2.py

@@ -5,7 +5,7 @@ Works on 5090 (32GB).
 
 import unittest
 
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import (
     DEFAULT_SMALL_EMBEDDING_MODEL_NAME_FOR_TEST,
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST_SCORE,
@@ -20,6 +20,7 @@ from sglang.test.test_utils import (
 )
 
 register_cuda_ci(est_time=900, suite="stage-b-test-large-1-gpu-performance")
+register_amd_ci(est_time=900, suite="stage-b-test-large-1-gpu-performance-amd")
 
 
 class TestBenchServing1GPUPart2(CustomTestCase):

test/registered/perf/test_bench_serving_2gpu.py

@@ -4,7 +4,7 @@ Performance tests for 2-GPU that need large GPUs (H200 80GB) - MoE and Pipeline
 
 import unittest
 
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import (
     DEFAULT_MOE_MODEL_NAME_FOR_TEST,
     CustomTestCase,
@@ -15,6 +15,7 @@ from sglang.test.test_utils import (
 )
 
 register_cuda_ci(est_time=600, suite="stage-b-test-large-2-gpu-performance")
+register_amd_ci(est_time=600, suite="stage-b-test-large-2-gpu-performance-amd")
 
 
 class TestBenchServing2GPU(CustomTestCase):

test/registered/perf/test_vlm_perf_5090.py

@@ -4,7 +4,7 @@ VLM Performance tests that work on 5090 (32GB) - VLM offline throughput and onli
 
 import unittest
 
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import (
     DEFAULT_SMALL_VLM_MODEL_NAME_FOR_TEST,
     CustomTestCase,
@@ -14,6 +14,7 @@ from sglang.test.test_utils import (
 )
 
 register_cuda_ci(est_time=600, suite="stage-b-test-small-1-gpu-performance")
+register_amd_ci(est_time=500, suite="stage-b-test-small-1-gpu-performance-amd")
 
 
 class TestVLMPerf5090(CustomTestCase):