[CI/Build] Reduce LoRA model tests. (#42266)

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
This commit is contained in:
Jee Jee Li
2026-05-11 14:49:08 +08:00
committed by GitHub
parent 581b5e9afc
commit 05d610e5cd
11 changed files with 50 additions and 31 deletions
+7
View File
@@ -11,6 +11,7 @@ from safetensors.torch import load_file, save_file
import vllm
from vllm.lora.request import LoRARequest
from vllm.platforms import current_platform
from ..utils import multi_gpu_test
@@ -110,6 +111,9 @@ def generate_and_test(
)
@pytest.mark.skipif(
current_platform.is_cuda_alike(), reason="Skipping to avoid redundant model tests"
)
def test_olmoe_lora(olmoe_lora_files, maybe_enable_lora_dual_stream):
# We enable enforce_eager=True here to reduce VRAM usage for lora-test CI,
# Otherwise, the lora-test will fail due to CUDA OOM.
@@ -178,6 +182,9 @@ def test_olmoe_lora_mixed_random(
assert outputs[0].outputs[0].text.strip().startswith(EXPECTED_LORA_OUTPUT[0])
@pytest.mark.skipif(
current_platform.is_cuda_alike(), reason="Skipping to avoid redundant model tests"
)
@pytest.mark.parametrize("fully_sharded_loras", [False, True])
@multi_gpu_test(num_gpus=2)
def test_olmoe_lora_tp2(olmoe_lora_files, fully_sharded_loras):