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

Unify GSM8K eval path to Chat API for regression CI readiness (#21667)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-04-01 17:12:19 -07:00
committed by GitHub
parent 1081a25983
commit d7256eb69a
79 changed files with 1349 additions and 1359 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/spec/eagle/test_deepseek_v3_fp4_mtp_small.py
+11 -12
View File
@@ -6,7 +6,7 @@ import requests
from sglang.srt.environ import envs
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.run_eval import run_eval
from sglang.test.send_one import BenchArgs, send_one_prompt
from sglang.test.test_utils import (
DEFAULT_URL_FOR_TEST,
@@ -18,7 +18,6 @@ from sglang.test.test_utils import (
register_cuda_ci(est_time=900, suite="stage-b-test-4-gpu-b200")
FULL_DEEPSEEK_V3_FP4_MODEL_PATH = "nvidia/DeepSeek-V3-0324-FP4"
SERVER_LAUNCH_TIMEOUT = 1200
@@ -74,15 +73,15 @@ class TestDeepseekV3FP4MTP(CustomTestCase):
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]),
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval_few_shot_gsm8k(args)
metrics = run_eval(args)
print(f"{metrics=}")
server_info = requests.get(self.base_url + "/server_info").json()
@@ -94,11 +93,11 @@ class TestDeepseekV3FP4MTP(CustomTestCase):
if is_in_ci():
write_github_step_summary(
f"### test_gsm8k (deepseek-v3-fp4 mtp)\n"
f'{metrics["accuracy"]=:.3f}\n'
f'{metrics["score"]=:.3f}\n'
f"{avg_spec_accept_length=:.2f}\n"
)
self.assertGreater(metrics["accuracy"], 0.94)
self.assertGreater(metrics["score"], 0.94)
self.assertGreater(avg_spec_accept_length, 2.7)
def test_bs_1_speed(self):
test/registered/spec/eagle/test_eagle_dp_attention.py
+12 -12
View File
@@ -5,7 +5,7 @@ import requests
from sglang.srt.environ import envs
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.run_eval import run_eval
from sglang.test.send_one import BenchArgs, send_one_prompt
from sglang.test.test_utils import (
DEFAULT_DRAFT_MODEL_EAGLE_DP_ATTN,
@@ -76,15 +76,15 @@ class TestEAGLE3EngineDPAttention(CustomTestCase):
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]),
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval_few_shot_gsm8k(args)
metrics = run_eval(args)
print(f"{metrics=}")
server_info = requests.get(self.base_url + "/get_server_info")
@@ -104,14 +104,14 @@ class TestEAGLE3EngineDPAttention(CustomTestCase):
if is_in_ci():
write_github_step_summary(
f"### test_gsm8k (EAGLE3 DP Attention)\n"
f'{metrics["accuracy"]=:.3f}\n'
f'{metrics["score"]=:.3f}\n'
f"{avg_spec_accept_length=:.2f}\n"
)
if is_in_amd_ci():
# AMD triton backend produces slightly lower accuracy than FA3 on NVIDIA
self.assertGreater(metrics["accuracy"], 0.88)
self.assertGreater(metrics["score"], 0.88)
else:
self.assertGreater(metrics["accuracy"], 0.91)
self.assertGreater(metrics["score"], 0.91)
if avg_spec_accept_length is not None:
if is_in_amd_ci():
# AMD triton backend produces slightly lower accept length than FA3 on NVIDIA
test/registered/spec/eagle/test_eagle_infer_b.py
+18 -18
View File
@@ -12,13 +12,13 @@ import requests
from sglang.srt.environ import envs
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval as run_gsm8k_eval
from sglang.test.kits.abort_timeout_kit import (
AbortAllMixin,
RunningTimeoutTwoWaveMixin,
WaitingTimeoutMixin,
)
from sglang.test.kits.radix_cache_server_kit import run_radix_attention_test
from sglang.test.run_eval import run_eval
from sglang.test.server_fixtures.eagle_fixture import EagleServerBase
from sglang.test.test_utils import DEFAULT_TARGET_MODEL_EAGLE, run_logprob_check
@@ -48,18 +48,18 @@ class TestEAGLEServerBasic(EagleServerBase):
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]),
base_url=self.base_url,
model=self.target_model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_gsm8k_eval(args)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreater(metrics["accuracy"], 0.20)
self.assertGreater(metrics["score"], 0.20)
server_info = requests.get(self.base_url + "/server_info").json()
avg_spec_accept_length = server_info["internal_states"][0][
@@ -103,16 +103,16 @@ class TestEAGLEServerAdditional(TestEAGLEServerBasic):
requests.get(self.base_url + "/flush_cache")
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=1,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
base_url=self.base_url,
model=self.target_model,
eval_name="gsm8k",
api="completion",
max_tokens=1,
num_examples=200,
num_threads=128,
)
metrics = run_gsm8k_eval(args)
metrics = run_eval(args)
self.assertGreater(metrics["output_throughput"], 50)
def test_logprob_start_len(self):
test/registered/spec/eagle/test_eagle_infer_beta.py
+10 -10
View File
@@ -7,9 +7,9 @@ import requests
from sglang.srt.environ import envs
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval
from sglang.test.kits.matched_stop_kit import MatchedStopMixin
from sglang.test.kits.radix_cache_server_kit import run_radix_attention_test
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_DRAFT_MODEL_EAGLE,
DEFAULT_TARGET_MODEL_EAGLE,
@@ -86,19 +86,19 @@ class TestEagleServerBase(CustomTestCase, MatchedStopMixin):
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=1000,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=1000,
num_threads=128,
)
metrics = run_eval(args)
print(f"TestEagleLargeBS -- {metrics=}")
self.assertGreater(
metrics["accuracy"], 0.23
) # 0.3333 for 60 questions; 0.234 for 1319 questions
metrics["score"], 0.22
) # ~0.227 for 1000 questions via /v1/completions
assert self.process.poll() is None
def test_logprob_spec_v2_match(self):
test/registered/spec/eagle/test_eagle_infer_beta_dp_attention.py
+12 -12
View File
@@ -6,7 +6,7 @@ import requests
from sglang.srt.environ import envs
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST_MLA,
DEFAULT_MODEL_NAME_FOR_TEST_MLA_NEXTN,
@@ -20,19 +20,19 @@ from sglang.test.test_utils import (
register_cuda_ci(est_time=300, suite="stage-c-test-4-gpu-b200")
def test_gsm8k(base_url: str):
def test_gsm8k(base_url: str, model: str):
requests.get(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(base_url.split(":")[-1]),
base_url=base_url,
model=model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval_few_shot_gsm8k(args)
metrics = run_eval(args)
server_info = requests.get(base_url + "/get_server_info")
avg_spec_accept_length = server_info.json()["internal_states"][0][
"avg_spec_accept_length"
@@ -84,8 +84,8 @@ class TestEagleDPAttnServerSmall(CustomTestCase):
kill_process_tree(cls.process.pid)
def test_a_gsm8k(self):
metrics, avg_spec_accept_length = test_gsm8k(self.base_url)
self.assertGreater(metrics["accuracy"], 0.62)
metrics, avg_spec_accept_length = test_gsm8k(self.base_url, self.model)
self.assertGreater(metrics["score"], 0.62)
self.assertGreater(avg_spec_accept_length, 2.7)
test/registered/spec/eagle/test_eagle_infer_beta_dp_attention_large.py
+13 -13
View File
@@ -6,7 +6,7 @@ import requests
from sglang.srt.environ import envs
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_DEEPSEEK_NVFP4_MODEL_FOR_TEST,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -21,19 +21,19 @@ from sglang.test.test_utils import (
register_cuda_ci(est_time=600, suite="nightly-8-gpu-b200", nightly=True)
def test_gsm8k(base_url: str):
def test_gsm8k(base_url: str, model: str):
requests.get(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(base_url.split(":")[-1]),
base_url=base_url,
model=model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval_few_shot_gsm8k(args)
metrics = run_eval(args)
server_info = requests.get(base_url + "/server_info").json()
avg_spec_accept_length = server_info["internal_states"][0]["avg_spec_accept_length"]
@@ -92,14 +92,14 @@ class TestEagleDPAttnServerLarge(CustomTestCase):
kill_process_tree(cls.process.pid)
def test_a_gsm8k(self):
metrics, avg_spec_accept_length = test_gsm8k(self.base_url)
metrics, avg_spec_accept_length = test_gsm8k(self.base_url, self.model)
self.assertGreater(metrics["accuracy"], 0.94)
self.assertGreater(metrics["score"], 0.94)
self.assertGreater(avg_spec_accept_length, 2.7)
if is_in_ci():
write_github_step_summary(
f"### test_gsm8k (deepseek-v3-fp4 mtp)\n"
f'{metrics["accuracy"]=:.3f}\n'
f'{metrics["score"]=:.3f}\n'
f"{avg_spec_accept_length=:.2f}\n"
)
test/registered/spec/test_standalone_speculative_decoding.py
+21 -18
View File
@@ -7,7 +7,7 @@ import requests
from sglang.srt.environ import envs
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_DRAFT_MODEL_STANDALONE,
DEFAULT_TARGET_MODEL_STANDALONE,
@@ -22,7 +22,6 @@ register_cuda_ci(est_time=308, suite="stage-b-test-1-gpu-large")
GSM_DATASET_PATH = None
# Default server arguments shared across all tests
DEFAULT_SERVER_ARGS = [
"--trust-remote-code",
@@ -97,19 +96,21 @@ class TestStandaloneSpeculativeDecodingBase(CustomTestCase):
requests.get(self.base_url + "/flush_cache")
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=100,
num_threads=128,
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,
gsm8k_data_path=GSM_DATASET_PATH,
)
metrics = run_eval_few_shot_gsm8k(args)
metrics = run_eval(args)
print(f"{metrics=}")
# Use the appropriate metric key based on the test class
metric_key = "accuracy"
metric_key = "score"
self.assertGreater(metrics[metric_key], self.accuracy_threshold)
server_info = requests.get(self.base_url + "/get_server_info")
@@ -158,19 +159,21 @@ class TestStandaloneV2SpeculativeDecodingBase(CustomTestCase):
requests.get(self.base_url + "/flush_cache")
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=100,
num_threads=128,
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,
gsm8k_data_path=GSM_DATASET_PATH,
)
metrics = run_eval_few_shot_gsm8k(args)
metrics = run_eval(args)
print(f"{metrics=}")
# Use the appropriate metric key based on the test class
metric_key = "accuracy"
metric_key = "score"
self.assertGreater(metrics[metric_key], self.accuracy_threshold)
server_info = requests.get(self.base_url + "/get_server_info")