chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

Tiny apply gsm8k mixin to ngram test (#15606)

Browse Source
This commit is contained in:
Liangsheng Yin
2025-12-24 01:30:26 +08:00
committed by GitHub
parent 5f3a47d8a7
commit bd572360f3
6 changed files with 31 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/srt/models/test_ministral3_models.py
+1 -1
View File
@@ -9,7 +9,7 @@ MODEL = "mistralai/Ministral-3-3B-Instruct-2512"
class TestMinistral3TextOnly(GSM8KMixin, DefaultServerBase):
accuracy = 0.6
gsm8k_accuracy_thres = 0.6
model = MODEL
other_args = ["--trust-remote-code"]
test/srt/models/test_nvidia_nemotron_nano_v2.py
+5 -5
View File
@@ -7,19 +7,19 @@ from sglang.test.server_fixtures.default_fixture import DefaultServerBase
class TestNvidiaNemotronNanoV2BF16(GSM8KMixin, DefaultServerBase):
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
accuracy = 0.87
gsm8k_accuracy_thres = 0.87
other_args = ["--max-mamba-cache-size", "256"]
class TestNvidiaNemotronNanoV2FP8(GSM8KMixin, DefaultServerBase):
accuracy = 0.87
gsm8k_accuracy_thres = 0.87
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8"
other_args = ["--max-mamba-cache-size", "256"]
@unittest.skipIf(not is_blackwell(), "NVFP4 only supported on blackwell")
class TestNvidiaNemotronNanoV2NVFP4(GSM8KMixin, DefaultServerBase):
accuracy = 0.855
gsm8k_accuracy_thres = 0.855
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2-NVFP4"
other_args = ["--max-mamba-cache-size", "256"]
@@ -29,7 +29,7 @@ class TestNvidiaNemotronNanoV2NVFP4(GSM8KMixin, DefaultServerBase):
"with different hidden sizes (Nemotron-9B: 4480, Llama-3.2-1B: 2048)"
)
class TestNvidiaNemotronNanoV2SpeculativeDecoding(GSM8KMixin, DefaultServerBase):
accuracy = 0.87
gsm8k_accuracy_thres = 0.87
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
other_args = [
"--speculative-algorithm",
@@ -60,7 +60,7 @@ class TestNvidiaNemotronNanoV2SpeculativeDecoding(GSM8KMixin, DefaultServerBase)
class TestNvidiaNemotronNanoV2SpeculativeDecodingBF16Cache(
GSM8KMixin, DefaultServerBase
):
accuracy = 0.87
gsm8k_accuracy_thres = 0.87
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
other_args = [
"--speculative-algorithm",
test/srt/models/test_nvidia_nemotron_nano_v2_vl.py
+1 -1
View File
@@ -9,7 +9,7 @@ MODEL = "nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16"
class TestNvidiaNemotronNanoV2VLTextOnly(GSM8KMixin, DefaultServerBase):
accuracy = 0.87
gsm8k_accuracy_thres = 0.87
model = MODEL
other_args = ["--max-mamba-cache-size", "256", "--trust-remote-code"]
test/srt/run_suite.py
-1
View File
@@ -79,7 +79,6 @@ suites = {
TestFile("test_model_hooks.py", 6),
TestFile("test_modelopt_loader.py", 11),
TestFile("test_multi_tokenizer.py", 230),
TestFile("test_ngram_speculative_decoding.py", 177),
TestFile("test_no_chunked_prefill.py", 108),
TestFile("test_no_overlap_scheduler.py", 217),
TestFile("test_original_logprobs.py", 41),
test/srt/test_ngram_speculative_decoding.py
-117
View File
@@ -1,117 +0,0 @@
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.environ import envs
from sglang.srt.utils import kill_process_tree
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_NGRAM_SPECULATIVE_TARGET_MODEL_FOR_TEST,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
GSM_DATASET_PATH = None
# Default server arguments shared across all tests
DEFAULT_SERVER_ARGS = [
"--trust-remote-code",
"--cuda-graph-max-bs",
"8",
"--speculative-algorithm",
"NGRAM",
"--speculative-num-draft-tokens",
"16",
"--mem-fraction-static",
0.8,
]
class TestNgramSpeculativeDecodingBase(CustomTestCase):
model = DEFAULT_NGRAM_SPECULATIVE_TARGET_MODEL_FOR_TEST
base_url = DEFAULT_URL_FOR_TEST
accuracy_threshold = 0.79 # derived tests need to override this
spec_decode_threshold = 1.8 # derived spec decoding tests need to override this
@classmethod
def get_server_args(cls):
"""Return the arguments for the server launch. Override in subclasses."""
return DEFAULT_SERVER_ARGS + ["--attention-backend", "fa3"]
@classmethod
def setUpClass(cls):
# disable deep gemm precompile to make launch server faster
# please don't do this if you want to make your inference workload faster
envs.SGLANG_JIT_DEEPGEMM_PRECOMPILE.set(False)
envs.SGLANG_ENABLE_JIT_DEEPGEMM.set(False)
model = cls.model
cls.process = popen_launch_server(
model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=cls.get_server_args(),
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
requests.get(self.base_url + "/flush_cache")
args = SimpleNamespace(
num_shots=4,
num_questions=100,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
data_path=GSM_DATASET_PATH,
)
metrics = run_eval_few_shot_gsm8k(args)
print(f"{metrics=}")
# Use the appropriate metric key based on the test class
metric_key = "accuracy"
self.assertGreater(metrics[metric_key], self.accuracy_threshold)
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=}")
self.assertGreater(avg_spec_accept_length, self.spec_decode_threshold)
class TestNgramSpeculativeDecodingTriton(TestNgramSpeculativeDecodingBase):
@classmethod
def get_server_args(cls):
return DEFAULT_SERVER_ARGS + ["--attention-backend", "triton"]
class TestNgramSpeculativeDecodingFlashinfer(TestNgramSpeculativeDecodingBase):
@classmethod
def get_server_args(cls):
return DEFAULT_SERVER_ARGS + ["--attention-backend", "flashinfer"]
class TestNgramSpeculativeDecodingPaged(TestNgramSpeculativeDecodingBase):
@classmethod
def get_server_args(cls):
return DEFAULT_SERVER_ARGS + [
"--attention-backend",
"flashinfer",
"--page-size",
"64",
]
if __name__ == "__main__":
unittest.main()