Support page size > 1 + eagle (#4908)

This commit is contained in:
Lianmin Zheng
2025-03-30 00:46:23 -07:00
committed by GitHub
parent 5ec5eaf760
commit b26bc86b36
16 changed files with 374 additions and 71 deletions

View File

@@ -76,11 +76,14 @@ def is_in_ci():
if is_in_ci():
DEFAULT_PORT_FOR_SRT_TEST_RUNNER = 5157
DEFAULT_URL_FOR_TEST = "http://127.0.0.1:6157"
DEFAULT_PORT_FOR_SRT_TEST_RUNNER = (
5000 + int(os.environ.get("CUDA_VISIBLE_DEVICES", "0")[0]) * 100
)
else:
DEFAULT_PORT_FOR_SRT_TEST_RUNNER = 1157
DEFAULT_URL_FOR_TEST = "http://127.0.0.1:2157"
DEFAULT_PORT_FOR_SRT_TEST_RUNNER = (
7000 + int(os.environ.get("CUDA_VISIBLE_DEVICES", "0")[0]) * 100
)
DEFAULT_URL_FOR_TEST = f"http://127.0.0.1:{DEFAULT_PORT_FOR_SRT_TEST_RUNNER + 1000}"
def call_generate_lightllm(prompt, temperature, max_tokens, stop=None, url=None):
@@ -1009,6 +1012,9 @@ def run_logprob_check(self: unittest.TestCase, arg: Tuple):
class CustomTestCase(unittest.TestCase):
pass
"""
def _callTestMethod(self, method):
max_retry = int(
os.environ.get("SGLANG_TEST_MAX_RETRY", "2" if is_in_ci() else "0")
@@ -1017,3 +1023,4 @@ class CustomTestCase(unittest.TestCase):
lambda: super(CustomTestCase, self)._callTestMethod(method),
max_retry=max_retry,
)
"""