chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 15 Packages Projects Releases Wiki Activity

enhance LoRA tests and fix base model LoRA eviction in Scheduler (#16333)

Browse Source
This commit is contained in:
Glen Liu
2026-01-10 16:49:00 +08:00
committed by GitHub
parent 675acecec6
commit 6327dff242
6 changed files with 144 additions and 117 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/manual/lora/test_lora_spec_decoding.py
+1 -19
View File
@@ -17,29 +17,11 @@ import unittest
from sglang.test.lora_utils import (
CI_MULTI_LORA_MODELS,
LoRAAdaptor,
LoRAModelCase,
LORA_MODELS_QWEN3,
run_lora_multiple_batch_on_model_cases,
)
from sglang.test.test_utils import CustomTestCase
LORA_MODELS_QWEN3 = [
LoRAModelCase(
base="Qwen/Qwen3-4B",
adaptors=[
LoRAAdaptor(
name="nissenj/Qwen3-4B-lora-v2",
prefill_tolerance=3e-1,
),
LoRAAdaptor(
name="y9760210/Qwen3-4B-lora_model",
prefill_tolerance=3e-1,
),
],
max_loras_per_batch=2,
),
]
class TestLoRASpecDecoding(CustomTestCase):
def test_qwen(self):
test/registered/lora/test_lora.py
-54
View File
@@ -1,54 +0,0 @@
# Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
import multiprocessing as mp
import os
import unittest
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
from sglang.test.lora_utils import (
ALL_OTHER_MULTI_LORA_MODELS,
CI_MULTI_LORA_MODELS,
run_lora_multiple_batch_on_model_cases,
)
from sglang.test.test_utils import CustomTestCase, is_in_ci
register_cuda_ci(est_time=82, suite="stage-b-test-small-1-gpu")
register_amd_ci(est_time=82, suite="stage-b-test-small-1-gpu-amd")
class TestLoRA(CustomTestCase):
def test_ci_lora_models(self):
run_lora_multiple_batch_on_model_cases(CI_MULTI_LORA_MODELS)
def test_all_lora_models(self):
if is_in_ci():
return
filtered_models = []
for model_case in ALL_OTHER_MULTI_LORA_MODELS:
if "ONLY_RUN" in os.environ and os.environ["ONLY_RUN"] != model_case.base:
continue
filtered_models.append(model_case)
run_lora_multiple_batch_on_model_cases(filtered_models)
if __name__ == "__main__":
try:
mp.set_start_method("spawn")
except RuntimeError:
pass
unittest.main(warnings="ignore")
test/registered/lora/test_lora_qwen3.py
+1 -19
View File
@@ -17,8 +17,7 @@ import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.lora_utils import (
LoRAAdaptor,
LoRAModelCase,
LORA_MODELS_QWEN3,
run_lora_multiple_batch_on_model_cases,
)
@@ -26,23 +25,6 @@ register_cuda_ci(est_time=97, suite="nightly-1-gpu", nightly=True)
from sglang.test.test_utils import CustomTestCase
LORA_MODELS_QWEN3 = [
LoRAModelCase(
base="Qwen/Qwen3-4B",
adaptors=[
LoRAAdaptor(
name="nissenj/Qwen3-4B-lora-v2",
prefill_tolerance=3e-1,
),
LoRAAdaptor(
name="y9760210/Qwen3-4B-lora_model",
prefill_tolerance=3e-1,
),
],
max_loras_per_batch=2,
),
]
class TestLoRAQwen3(CustomTestCase):
def test_ci_lora_models(self):
test/registered/lora/test_multi_lora_backend.py
+7 -17
View File
@@ -20,30 +20,20 @@ from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
from sglang.test.lora_utils import (
ALL_OTHER_MULTI_LORA_MODELS,
CI_MULTI_LORA_MODELS,
run_lora_batch_splitting_equivalence_test,
run_lora_multiple_batch_on_model_cases,
)
from sglang.test.test_utils import CustomTestCase, is_in_ci
register_cuda_ci(est_time=60, suite="stage-b-test-small-1-gpu")
register_amd_ci(est_time=60, suite="stage-b-test-small-1-gpu-amd")
# All prompts are used at once in a batch.
PROMPTS = [
"AI is a field of computer science focused on",
"""
### Instruction:
Tell me about llamas and alpacas
### Response:
Llamas are large, long-necked animals with a woolly coat. They have two toes on each foot instead of three like other camelids.
### Question:
What do you know about llamas?
### Answer:
""",
]
register_cuda_ci(est_time=100, suite="stage-b-test-small-1-gpu")
register_amd_ci(est_time=100, suite="stage-b-test-small-1-gpu-amd")
class TestMultiLoRABackend(CustomTestCase):
def test_ci_lora_models(self):
def test_ci_lora_models_batch_splitting(self):
run_lora_batch_splitting_equivalence_test(CI_MULTI_LORA_MODELS)
def test_ci_lora_models_multi_batch(self):
run_lora_multiple_batch_on_model_cases(CI_MULTI_LORA_MODELS)
def test_all_lora_models(self):