EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Simplify tests & Fix trtllm custom allreduce registration (#4252)

Browse Source
This commit is contained in:
Lianmin Zheng
2025-03-10 01:24:22 -07:00
committed by GitHub
parent 007f8b3dc2
commit aa957102a9
13 changed files with 30 additions and 211 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
test/srt/run_suite.py
View File

@@ -20,30 +20,33 @@ suites = {
TestFile("models/test_generation_models.py", 103),
TestFile("models/test_qwen_models.py", 82),
TestFile("models/test_reward_models.py", 83),
TestFile("test_gptqmodel_dynamic.py", 72),
TestFile("models/test_gme_qwen_models.py", 45),
TestFile("test_abort.py", 51),
TestFile("test_block_int8.py", 22),
TestFile("test_chunked_prefill.py", 336),
TestFile("test_custom_allreduce.py", 1),
TestFile("test_double_sparsity.py", 50),
TestFile("test_eagle_infer.py", 447),
TestFile("test_ebnf_constrained.py"),
TestFile("test_fp8_kernel.py", 2),
TestFile("test_embedding_openai_server.py", 36),
TestFile("test_eval_accuracy_mini.py", 63),
TestFile("test_gguf.py", 78),
TestFile("test_gptqmodel_dynamic.py", 72),
TestFile("test_hidden_states.py", 55),
TestFile("test_int8_kernel.py", 1),
TestFile("test_input_embeddings.py", 38),
TestFile("test_json_constrained.py", 98),
TestFile("test_large_max_new_tokens.py", 41),
TestFile("test_metrics.py", 32),
TestFile("test_mla.py", 92),
TestFile("test_mla_deepseek_v3.py", 221),
TestFile("test_mla_flashinfer.py", 395),
TestFile("test_mla_fp8.py", 93),
TestFile("test_json_constrained.py", 98),
TestFile("test_large_max_new_tokens.py", 41),
TestFile("test_metrics.py", 32),
TestFile("test_no_chunked_prefill.py", 126),
TestFile("test_no_overlap_scheduler.py", 262),
TestFile("test_openai_server.py", 124),
TestFile("test_penalty.py", 41),
TestFile("test_pytorch_sampling_backend.py", 66),
TestFile("test_radix_attention.py", 167),
TestFile("test_reasoning_content.py", 89),
TestFile("test_regex_constrained.py", 64),
TestFile("test_release_memory_occupation.py", 44),
TestFile("test_request_length_validation.py", 31),
@@ -58,7 +61,6 @@ suites = {
TestFile("test_torchao.py", 70),
TestFile("test_triton_attention_kernels.py", 4),
TestFile("test_triton_attention_backend.py", 134),
TestFile("test_hidden_states.py", 55),
TestFile("test_update_weights_from_disk.py", 114),
TestFile("test_update_weights_from_tensor.py", 48),
TestFile("test_vertex_endpoint.py", 31),
@@ -66,10 +68,6 @@ suites = {
TestFile("test_vision_llm.py", 18.4),
TestFile("test_vision_openai_server.py", 344),
TestFile("test_w8a8_quantization.py", 46),
TestFile("test_fp8_kernel.py", 2),
TestFile("test_block_int8.py", 22),
TestFile("test_int8_kernel.py", 1),
TestFile("test_reasoning_content.py", 89),
],
"nightly": [
TestFile("test_nightly_gsm8k_eval.py"),

7
test/srt/test_bench_one_batch.py
View File

@@ -3,6 +3,7 @@ import unittest
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST,
DEFAULT_MOE_MODEL_NAME_FOR_TEST,
get_bool_env_var,
is_in_ci,
run_bench_one_batch,
write_github_step_summary,
@@ -27,9 +28,13 @@ class TestBenchOneBatch(unittest.TestCase):
DEFAULT_MOE_MODEL_NAME_FOR_TEST, ["--tp", "2", "--cuda-graph-max-bs", "2"]
)
use_vllm_custom_allreduce = get_bool_env_var(
"USE_VLLM_CUSTOM_ALLREDUCE", default="true"
)
if is_in_ci():
write_github_step_summary(
f"### test_moe_tp2_bs1\n"
f"### test_moe_tp2_bs1 ({use_vllm_custom_allreduce=})\n"
f"output_throughput : {output_throughput:.2f} token/s\n"
)
self.assertGreater(output_throughput, 124)

68
test/srt/test_eval_accuracy_large_chunked_prefill.py
View File

@@ -1,68 +0,0 @@
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
popen_launch_server,
)
class TestEvalAccuracyLargeChunkedPrefill(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_MODEL_NAME_FOR_TEST
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=["--log-level-http", "warning", "--chunked-prefill-size", "256"],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_mmlu(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="mmlu",
num_examples=3000,
num_threads=1024,
)
metrics = run_eval(args)
assert metrics["score"] >= 0.705, f"{metrics}"
def test_human_eval(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="humaneval",
num_examples=None,
num_threads=1024,
)
metrics = run_eval(args)
assert metrics["score"] >= 0.64, f"{metrics}"
def test_mgsm_en(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="mgsm_en",
num_examples=None,
num_threads=1024,
)
metrics = run_eval(args)
assert metrics["score"] >= 0.84, f"{metrics}"
if __name__ == "__main__":
unittest.main()

74
test/srt/test_eval_accuracy_large_mixed_chunked_prefill.py
View File

@@ -1,74 +0,0 @@
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
popen_launch_server,
)
class TestEvalAccuracyLargeChunkedPrefill(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_MODEL_NAME_FOR_TEST
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--log-level-http",
"warning",
"--chunked-prefill-size",
"256",
"--enable-mixed-chunk",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_mmlu(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="mmlu",
num_examples=3000,
num_threads=1024,
)
metrics = run_eval(args)
assert metrics["score"] >= 0.705, f"{metrics}"
def test_human_eval(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="humaneval",
num_examples=None,
num_threads=1024,
)
metrics = run_eval(args)
assert metrics["score"] >= 0.64, f"{metrics}"
def test_mgsm_en(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="mgsm_en",
num_examples=None,
num_threads=1024,
)
metrics = run_eval(args)
assert metrics["score"] >= 0.84, f"{metrics}"
if __name__ == "__main__":
unittest.main()

42
test/srt/test_eval_accuracy_mini.py
View File

@@ -1,42 +0,0 @@
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
popen_launch_server,
)
class TestEvalAccuracyMini(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_MODEL_NAME_FOR_TEST
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model, cls.base_url, timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_mmlu(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="mmlu",
num_examples=64,
num_threads=32,
temperature=0.1,
)
metrics = run_eval(args)
self.assertGreaterEqual(metrics["score"], 0.65)
if __name__ == "__main__":
unittest.main()

1
test/srt/test_gptqmodel_dynamic.py
View File

@@ -129,6 +129,7 @@ class TestGPTQModelDynamic(unittest.TestCase):
"text": "The capital of France is",
"sampling_params": {
"max_new_tokens": max_new_tokens,
"temperature": 0.001,
},
},
)