chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[AMD CI] Migrate and Add More Testcases (#17116)

Browse Source 
Co-authored-by: yctseng0211 <yctseng@amd.com>
This commit is contained in:
Bingxu Chen
2026-01-20 00:07:39 +08:00
committed by GitHub
parent 20b0523eca
commit 2ea02f0642
19 changed files with 311 additions and 67 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
test/registered/amd/test_deepseek_r1_mxfp4_8gpu.py
View File

@@ -19,7 +19,7 @@ from sglang.test.test_utils import (
register_amd_ci(est_time=3600, suite="stage-c-test-large-8-gpu-amd-mi35x")
DEEPSEEK_R1_MODEL_PATH = "amd/DeepSeek-R1-MXFP4-Preview"
SERVER_LAUNCH_TIMEOUT = 1200
SERVER_LAUNCH_TIMEOUT = 1800
class TestDeepseekR1MXFP4(CustomTestCase):

84
test/registered/amd/test_deepseek_v3_basic.py Normal file
View File

@@ -0,0 +1,84 @@
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_amd_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.send_one import BenchArgs, send_one_prompt
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
is_in_amd_ci,
is_in_ci,
popen_launch_server,
write_github_step_summary,
)
register_amd_ci(est_time=952, suite="stage-c-test-large-8-gpu-amd")
FULL_DEEPSEEK_V3_MODEL_PATH = "deepseek-ai/DeepSeek-V3-0324"
class TestDeepseekV3Basic(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = FULL_DEEPSEEK_V3_MODEL_PATH
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--trust-remote-code",
"--tp",
"8",
"--model-loader-extra-config",
'{"enable_multithread_load": true, "num_threads": 64}',
]
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 5,
other_args=other_args,
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_a_gsm8k(
self,
): # Append an "a" to make this test run first (alphabetically) to warm up the server
args = SimpleNamespace(
num_shots=8,
data_path=None,
num_questions=1400,
parallel=1400,
max_new_tokens=512,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval_few_shot_gsm8k(args)
print(f"{metrics=}")
if is_in_ci():
write_github_step_summary(
f"### test_gsm8k (deepseek-v3)\n" f'{metrics["accuracy"]=:.3f}\n'
)
self.assertGreater(metrics["accuracy"], 0.935)
def test_bs_1_speed(self):
args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
acc_length, speed = send_one_prompt(args)
print(f"{speed=:.2f}")
if is_in_ci():
write_github_step_summary(
f"### test_bs_1_speed (deepseek-v3)\n" f"{speed=:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(speed, 12)
else:
self.assertGreater(speed, 75)
if __name__ == "__main__":
unittest.main()

116
test/registered/amd/test_deepseek_v3_mtp.py Normal file
View File

@@ -0,0 +1,116 @@
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_amd_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.send_one import BenchArgs, send_one_prompt
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
is_in_amd_ci,
is_in_ci,
popen_launch_server,
write_github_step_summary,
)
register_amd_ci(est_time=980, suite="stage-c-test-large-8-gpu-amd")
FULL_DEEPSEEK_V3_MODEL_PATH = "deepseek-ai/DeepSeek-V3-0324"
class TestDeepseekV3MTP(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = FULL_DEEPSEEK_V3_MODEL_PATH
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--tp",
"8",
"--trust-remote-code",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--model-loader-extra-config",
'{"enable_multithread_load": true, "num_threads": 64}',
]
if not is_in_amd_ci():
other_args += ["--mem-frac", "0.7"]
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 5,
other_args=other_args,
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_a_gsm8k(
self,
): # Append an "a" to make this test run first (alphabetically) to warm up the server
requests.get(self.base_url + "/flush_cache")
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval_few_shot_gsm8k(args)
print(f"{metrics=}")
server_info = requests.get(self.base_url + "/get_server_info")
avg_spec_accept_length = server_info.json()["internal_states"][0][
"avg_spec_accept_length"
]
print(f"{avg_spec_accept_length=}")
if is_in_ci():
write_github_step_summary(
f"### test_gsm8k (deepseek-v3 mtp)\n"
f'{metrics["accuracy"]=:.3f}\n'
f"{avg_spec_accept_length=:.2f}\n"
)
self.assertGreater(metrics["accuracy"], 0.935)
if is_in_amd_ci():
self.assertGreater(avg_spec_accept_length, 2.8)
else:
self.assertGreater(avg_spec_accept_length, 2.9)
def test_bs_1_speed(self):
args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
acc_length, speed = send_one_prompt(args)
print(f"{acc_length=:.2f} {speed=:.2f}")
if is_in_ci():
write_github_step_summary(
f"### test_bs_1_speed (deepseek-v3 mtp)\n"
f"{acc_length=:.2f}\n"
f"{speed=:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(acc_length, 2.8)
else:
self.assertGreater(acc_length, 2.9)
if is_in_amd_ci():
self.assertGreater(speed, 15)
else:
self.assertGreater(speed, 130)
if __name__ == "__main__":
unittest.main()

2
test/registered/attention/test_wave_attention_kernels.py
View File

@@ -24,7 +24,7 @@ from sglang.srt.layers.attention.wave_ops.prefill_attention import (
from sglang.test.ci.ci_register import register_amd_ci
# Wave attention kernel unit tests (AMD only - requires wave_lang)
register_amd_ci(est_time=60, suite="stage-a-test-1")
register_amd_ci(est_time=60, suite="stage-a-test-1-amd")
class TestWaveAttention(unittest.TestCase):

6
test/registered/core/test_deterministic.py
View File

@@ -9,15 +9,18 @@ test into unit tests so that's easily reproducible in CI.
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_deterministic_utils import (
COMMON_SERVER_ARGS,
TestDeterministicBase,
)
from sglang.test.test_utils import is_in_amd_ci
register_cuda_ci(est_time=278, suite="stage-b-test-large-1-gpu")
register_amd_ci(est_time=278, suite="stage-b-test-small-1-gpu-amd")
@unittest.skipIf(is_in_amd_ci(), "Skip for AMD CI.")
class TestFlashinferDeterministic(TestDeterministicBase):
# Test with flashinfer attention backend
@classmethod
@@ -32,6 +35,7 @@ class TestFlashinferDeterministic(TestDeterministicBase):
return args
@unittest.skipIf(is_in_amd_ci(), "Skip for AMD CI.")
class TestFa3Deterministic(TestDeterministicBase):
# Test with fa3 attention backend
@classmethod

3
test/registered/hicache/test_hicache_storage_3fs_backend.py
View File

@@ -10,10 +10,11 @@ import unittest
from test_hicache_storage_file_backend import HiCacheStorageBaseMixin
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 CustomTestCase
register_cuda_ci(est_time=200, suite="stage-b-test-large-2-gpu")
register_amd_ci(est_time=300, suite="stage-b-test-large-2-gpu")
class HiCacheStorage3FSBackendBaseMixin(HiCacheStorageBaseMixin):

3
test/registered/hicache/test_hicache_storage_file_backend.py
View File

@@ -18,7 +18,7 @@ import requests
from sglang.bench_serving import get_tokenizer
from sglang.srt.utils import kill_process_tree
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.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_MLA_MODEL_NAME_FOR_TEST,
@@ -31,6 +31,7 @@ from sglang.test.test_utils import (
)
register_cuda_ci(est_time=200, suite="stage-b-test-large-2-gpu")
register_amd_ci(est_time=526, suite="stage-b-test-large-2-gpu-amd")
class HiCacheStorageBaseMixin:

2
test/registered/openai_server/basic/test_serving_rerank.py
View File

@@ -7,7 +7,7 @@ from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
# Keep consistent with other openai_server/basic unit tests.
register_cuda_ci(est_time=10, suite="stage-b-test-large-1-gpu")
register_amd_ci(est_time=10, suite="stage-b-test-small-1-gpu")
register_amd_ci(est_time=10, suite="stage-b-test-small-1-gpu-amd")
try:
from sglang.srt.entrypoints.openai.serving_rerank import (

2
test/registered/quant/test_awq_dequant.py
View File

@@ -19,7 +19,7 @@ from sglang.srt.layers.quantization.awq_triton import (
from sglang.test.ci.ci_register import register_amd_ci
from sglang.test.test_utils import CustomTestCase
register_amd_ci(est_time=2, suite="stage-a-test-1")
register_amd_ci(est_time=2, suite="stage-a-test-1-amd")
device = "cuda"

2
test/registered/quant/test_fused_rms_fp8_group_quant.py
View File

@@ -7,7 +7,7 @@ import torch.nn.functional as F
from sglang.test.ci.ci_register import register_amd_ci
from sglang.test.test_utils import CustomTestCase
register_amd_ci(est_time=10, suite="stage-a-test-1")
register_amd_ci(est_time=10, suite="stage-a-test-1-amd")
def _fp8_available() -> bool:

3
test/srt/test_int4fp8_moe.py → test/registered/quant/test_int4fp8_moe.py
View File

@@ -1,6 +1,7 @@
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_amd_ci
from sglang.test.few_shot_gsm8k import run_eval
from sglang.test.test_utils import (
DEFAULT_URL_FOR_TEST,
@@ -8,6 +9,8 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_amd_ci(est_time=313, suite="stage-b-test-small-1-gpu-amd")
class TestMixtralAccuracy(CustomTestCase):
@classmethod

2
test/registered/rl/test_lora_load_from_tensor.py
View File

@@ -11,7 +11,7 @@ from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
from sglang.test.test_utils import CustomTestCase
register_cuda_ci(est_time=90, suite="stage-b-test-small-1-gpu")
register_amd_ci(est_time=90, suite="stage-b-test-small-1-gpu")
register_amd_ci(est_time=90, suite="stage-b-test-small-1-gpu-amd")
MODEL_PATH = "Qwen/Qwen3-0.6B"
LORA_REPO = "charent/self_cognition_Alice"

3
test/srt/test_rope_rocm.py → test/registered/rotary/test_rope_rocm.py
View File

@@ -4,8 +4,11 @@ import torch
from sglang.srt.layers.rotary_embedding import RotaryEmbedding
from sglang.srt.utils import get_bool_env_var, is_hip
from sglang.test.ci.ci_register import register_amd_ci
from sglang.test.test_utils import CustomTestCase
register_amd_ci(est_time=3, suite="stage-b-test-small-1-gpu-amd")
torch.manual_seed(0)
_is_hip = is_hip()

2
test/registered/test_srt_backend.py
View File

@@ -20,7 +20,7 @@ from sglang.test.test_programs import (
from sglang.test.test_utils import DEFAULT_MODEL_NAME_FOR_TEST, CustomTestCase
register_cuda_ci(est_time=80, suite="stage-a-test-1")
register_amd_ci(est_time=120, suite="stage-a-test-1")
register_amd_ci(est_time=120, suite="stage-a-test-1-amd")
class TestSRTBackend(CustomTestCase):

3
test/srt/test_bench_typebaseddispatcher.py → test/registered/utils/test_bench_typebaseddispatcher.py
View File

@@ -1,8 +1,11 @@
import timeit
from typing import Any, Callable, List, Tuple, Type
from sglang.test.ci.ci_register import register_amd_ci
from sglang.utils import TypeBasedDispatcher
register_amd_ci(est_time=10, suite="stage-b-test-small-1-gpu-amd")
class TypeBasedDispatcherList:
def __init__(self, mapping: List[Tuple[Type, Callable]]):

3
test/srt/test_type_based_dispatcher.py → test/registered/utils/test_type_based_dispatcher.py
View File

@@ -8,8 +8,11 @@ import timeit
import unittest
from sglang.srt.managers.io_struct import SamplingParams
from sglang.test.ci.ci_register import register_amd_ci
from sglang.utils import TypeBasedDispatcher
register_amd_ci(est_time=10, suite="stage-b-test-small-1-gpu-amd")
class TestTypeBasedDispatcher(unittest.TestCase):
"""Unit tests for TypeBasedDispatcher e2e performance."""

2
test/run_suite.py
View File

@@ -19,7 +19,7 @@ HW_MAPPING = {
PER_COMMIT_SUITES = {
HWBackend.CPU: ["default", "stage-a-cpu-only"],
HWBackend.AMD: [
"stage-a-test-1",
"stage-a-test-1-amd",
"stage-b-test-small-1-gpu-amd",
"stage-b-test-small-1-gpu-amd-mi35x",
"stage-b-test-large-2-gpu-amd",

8
test/srt/run_suite.py
View File

@@ -80,10 +80,7 @@ suite_amd = {
# TestFile("lora/test_lora_backend.py", 99), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/13107
# TestFile("lora/test_lora_cuda_graph.py", 250), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/13107
# TestFile("lora/test_lora_qwen3.py", 97), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/13107
TestFile("test_bench_typebaseddispatcher.py", 10),
TestFile("test_rope_rocm.py", 3),
# TestFile("test_torch_compile_moe.py", 210), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/13107
TestFile("test_type_based_dispatcher.py", 10),
# Disabled temporarily
# TestFile("test_vlm_input_format.py", 300),
# TestFile("openai_server/features/test_openai_server_hidden_states.py", 240),
@@ -92,15 +89,10 @@ suite_amd = {
# TestFile("test_vision_chunked_prefill.py", 175), # Disabled temporarily and track in #7701
# TestFile("test_wave_attention_backend.py", 150), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/11127
# The time estimation for `test_int4fp8_moe.py` assumes `mistralai/Mixtral-8x7B-Instruct-v0.1` is already cached (running on 1xMI300X).
TestFile("test_int4fp8_moe.py", 313),
],
"per-commit-4-gpu-amd": [
TestFile("test_pp_single_node.py", 150),
],
"per-commit-8-gpu-amd": [
TestFile("test_deepseek_v3_basic.py", 275),
TestFile("test_deepseek_v3_mtp.py", 275),
],
# NOTE: AMD nightly suites (nightly-amd, nightly-amd-vlm, nightly-amd-8-gpu)
# have been migrated to test/registered/amd/nightly/ and are now managed
# by test/run_suite.py using the registry system.