[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
+2 -2
View File
@@ -460,7 +460,7 @@ steps:
- tests/lora - tests/lora
- vllm/platforms/rocm.py - vllm/platforms/rocm.py
commands: commands:
- pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_llm_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py --ignore=lora/test_qwen35_densemodel_lora.py - pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_qwen3_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py --ignore=lora/test_qwen35_densemodel_lora.py
#------------------------------------------------------ mi250 · model_executor -------------------------------------------------------# #------------------------------------------------------ mi250 · model_executor -------------------------------------------------------#
@@ -1760,7 +1760,7 @@ steps:
- export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True - export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
- pytest -v -s -x lora/test_chatglm3_tp.py - pytest -v -s -x lora/test_chatglm3_tp.py
- pytest -v -s -x lora/test_llama_tp.py - pytest -v -s -x lora/test_llama_tp.py
- pytest -v -s -x lora/test_llm_with_multi_loras.py - pytest -v -s -x lora/test_qwen3_with_multi_loras.py
- pytest -v -s -x lora/test_olmoe_tp.py - pytest -v -s -x lora/test_olmoe_tp.py
- pytest -v -s -x lora/test_gptoss_tp.py - pytest -v -s -x lora/test_gptoss_tp.py
- pytest -v -s -x lora/test_qwen35_densemodel_lora.py - pytest -v -s -x lora/test_qwen35_densemodel_lora.py
+2 -2
View File
@@ -9,7 +9,7 @@ steps:
- vllm/lora - vllm/lora
- tests/lora - tests/lora
commands: commands:
- pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_llm_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py --ignore=lora/test_qwen35_densemodel_lora.py - pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_qwen3_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py --ignore=lora/test_qwen35_densemodel_lora.py
parallelism: 4 parallelism: 4
@@ -31,7 +31,7 @@ steps:
# requires multi-GPU testing for validation. # requires multi-GPU testing for validation.
- pytest -v -s -x lora/test_chatglm3_tp.py - pytest -v -s -x lora/test_chatglm3_tp.py
- pytest -v -s -x lora/test_llama_tp.py - pytest -v -s -x lora/test_llama_tp.py
- pytest -v -s -x lora/test_llm_with_multi_loras.py - pytest -v -s -x lora/test_qwen3_with_multi_loras.py
- pytest -v -s -x lora/test_olmoe_tp.py - pytest -v -s -x lora/test_olmoe_tp.py
- pytest -v -s -x lora/test_gptoss_tp.py - pytest -v -s -x lora/test_gptoss_tp.py
- pytest -v -s -x lora/test_qwen35_densemodel_lora.py - pytest -v -s -x lora/test_qwen35_densemodel_lora.py
+9
View File
@@ -1,9 +1,12 @@
# SPDX-License-Identifier: Apache-2.0 # SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
import vllm import vllm
import vllm.config import vllm.config
from vllm.lora.request import LoRARequest from vllm.lora.request import LoRARequest
from vllm.platforms import current_platform
from ..utils import create_new_process_for_each_test, multi_gpu_test from ..utils import create_new_process_for_each_test, multi_gpu_test
@@ -50,6 +53,9 @@ def do_sample(llm: vllm.LLM, lora_path: str, lora_id: int) -> list[str]:
return generated_texts return generated_texts
@pytest.mark.skipif(
current_platform.is_cuda_alike(), reason="Skipping to avoid redundant model tests"
)
@create_new_process_for_each_test() @create_new_process_for_each_test()
def test_chatglm3_lora(chatglm3_lora_files): def test_chatglm3_lora(chatglm3_lora_files):
llm = vllm.LLM( llm = vllm.LLM(
@@ -70,6 +76,9 @@ def test_chatglm3_lora(chatglm3_lora_files):
assert output2[i] == EXPECTED_LORA_OUTPUT[i] assert output2[i] == EXPECTED_LORA_OUTPUT[i]
@pytest.mark.skipif(
current_platform.is_cuda_alike(), reason="Skipping to avoid redundant model tests"
)
@multi_gpu_test(num_gpus=4) @multi_gpu_test(num_gpus=4)
def test_chatglm3_lora_tp4(chatglm3_lora_files): def test_chatglm3_lora_tp4(chatglm3_lora_files):
llm = vllm.LLM( llm = vllm.LLM(
+10
View File
@@ -11,6 +11,7 @@ import pytest
from huggingface_hub import snapshot_download from huggingface_hub import snapshot_download
from vllm.lora.request import LoRARequest from vllm.lora.request import LoRARequest
from vllm.platforms import current_platform
from ..conftest import AudioTestAssets, VllmRunner from ..conftest import AudioTestAssets, VllmRunner
from ..utils import create_new_process_for_each_test from ..utils import create_new_process_for_each_test
@@ -76,6 +77,9 @@ def test_active_default_mm_lora(
) )
@pytest.mark.skipif(
current_platform.is_cuda_alike(), reason="Skipping to avoid redundant model tests"
)
@create_new_process_for_each_test() @create_new_process_for_each_test()
def test_inactive_default_mm_lora( def test_inactive_default_mm_lora(
vllm_runner: type[VllmRunner], vllm_runner: type[VllmRunner],
@@ -92,6 +96,9 @@ def test_inactive_default_mm_lora(
) )
@pytest.mark.skipif(
current_platform.is_cuda_alike(), reason="Skipping to avoid redundant model tests"
)
@create_new_process_for_each_test() @create_new_process_for_each_test()
def test_default_mm_lora_succeeds_with_redundant_lora_request( def test_default_mm_lora_succeeds_with_redundant_lora_request(
vllm_runner: type[VllmRunner], vllm_runner: type[VllmRunner],
@@ -107,6 +114,9 @@ def test_default_mm_lora_succeeds_with_redundant_lora_request(
) )
@pytest.mark.skipif(
current_platform.is_cuda_alike(), reason="Skipping to avoid redundant model tests"
)
@create_new_process_for_each_test() @create_new_process_for_each_test()
def test_default_mm_lora_fails_with_overridden_lora_request( def test_default_mm_lora_fails_with_overridden_lora_request(
vllm_runner: type[VllmRunner], vllm_runner: type[VllmRunner],
+4
View File
@@ -10,6 +10,7 @@ import vllm.config
from vllm import LLM from vllm import LLM
from vllm.lora.request import LoRARequest from vllm.lora.request import LoRARequest
from vllm.model_executor.model_loader.tensorizer import TensorizerConfig from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
from vllm.platforms import current_platform
from ..utils import VLLM_PATH, create_new_process_for_each_test, multi_gpu_test from ..utils import VLLM_PATH, create_new_process_for_each_test, multi_gpu_test
@@ -139,6 +140,9 @@ def test_llama_lora(llama32_lora_files, cudagraph_specialize_lora: bool):
generate_and_test(llm, llama32_lora_files) generate_and_test(llm, llama32_lora_files)
@pytest.mark.skipif(
current_platform.is_cuda_alike(), reason="Skipping to avoid redundant model tests"
)
@multi_gpu_test(num_gpus=4) @multi_gpu_test(num_gpus=4)
def test_llama_lora_tp4(llama32_lora_files): def test_llama_lora_tp4(llama32_lora_files):
llm = vllm.LLM( llm = vllm.LLM(
+3
View File
@@ -68,6 +68,9 @@ def do_sample(llm: vllm.LLM, lora_path: str, lora_id: int) -> list[str]:
return generated_texts return generated_texts
@pytest.mark.skipif(
current_platform.is_cuda_alike(), reason="Skipping to avoid redundant model tests"
)
def test_minicpmv_lora(minicpmv_lora_files): def test_minicpmv_lora(minicpmv_lora_files):
llm = vllm.LLM( llm = vllm.LLM(
MODEL_PATH, MODEL_PATH,
+7
View File
@@ -11,6 +11,7 @@ from safetensors.torch import load_file, save_file
import vllm import vllm
from vllm.lora.request import LoRARequest from vllm.lora.request import LoRARequest
from vllm.platforms import current_platform
from ..utils import multi_gpu_test 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): 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, # We enable enforce_eager=True here to reduce VRAM usage for lora-test CI,
# Otherwise, the lora-test will fail due to CUDA OOM. # 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]) 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]) @pytest.mark.parametrize("fully_sharded_loras", [False, True])
@multi_gpu_test(num_gpus=2) @multi_gpu_test(num_gpus=2)
def test_olmoe_lora_tp2(olmoe_lora_files, fully_sharded_loras): def test_olmoe_lora_tp2(olmoe_lora_files, fully_sharded_loras):
@@ -1,12 +1,14 @@
# SPDX-License-Identifier: Apache-2.0 # SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from transformers import AutoTokenizer from transformers import AutoTokenizer
import vllm import vllm
import vllm.config import vllm.config
from vllm.assets.image import ImageAsset from vllm.assets.image import ImageAsset
from vllm.lora.request import LoRARequest from vllm.lora.request import LoRARequest
from vllm.platforms import current_platform
from ..utils import create_new_process_for_each_test, multi_gpu_test from ..utils import create_new_process_for_each_test, multi_gpu_test
@@ -311,6 +313,9 @@ def _assert_qwen35_text_vl_and_mixed_lora(
) )
@pytest.mark.skipif(
current_platform.is_cuda_alike(), reason="Skipping to avoid redundant model tests"
)
@create_new_process_for_each_test() @create_new_process_for_each_test()
def test_qwen35_text_lora( def test_qwen35_text_lora(
qwen35_text_lora_files, qwen35_vl_lora_files, maybe_enable_lora_dual_stream qwen35_text_lora_files, qwen35_vl_lora_files, maybe_enable_lora_dual_stream
+8
View File
@@ -2,12 +2,14 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from dataclasses import dataclass from dataclasses import dataclass
import pytest
from packaging.version import Version from packaging.version import Version
from transformers import __version__ as TRANSFORMERS_VERSION from transformers import __version__ as TRANSFORMERS_VERSION
import vllm import vllm
from vllm.assets.image import ImageAsset from vllm.assets.image import ImageAsset
from vllm.lora.request import LoRARequest from vllm.lora.request import LoRARequest
from vllm.platforms import current_platform
from vllm.sampling_params import BeamSearchParams from vllm.sampling_params import BeamSearchParams
@@ -206,6 +208,9 @@ def test_qwen2vl_lora_beam_search(qwen2vl_lora_files):
) )
@pytest.mark.skipif(
current_platform.is_cuda_alike(), reason="Skipping to avoid redundant model tests"
)
def test_qwen25vl_lora(qwen25vl_lora_files): def test_qwen25vl_lora(qwen25vl_lora_files):
"""Test Qwen 2.5 VL model with LoRA""" """Test Qwen 2.5 VL model with LoRA"""
config = TestConfig(model_path=QWEN25VL_MODEL_PATH, lora_path=qwen25vl_lora_files) config = TestConfig(model_path=QWEN25VL_MODEL_PATH, lora_path=qwen25vl_lora_files)
@@ -216,6 +221,9 @@ def test_qwen25vl_lora(qwen25vl_lora_files):
tester.run_test(TEST_IMAGES, expected_outputs=EXPECTED_OUTPUTS, lora_id=lora_id) tester.run_test(TEST_IMAGES, expected_outputs=EXPECTED_OUTPUTS, lora_id=lora_id)
@pytest.mark.skipif(
current_platform.is_cuda_alike(), reason="Skipping to avoid redundant model tests"
)
def test_qwen25vl_vision_lora(qwen25vl_vision_lora_files): def test_qwen25vl_vision_lora(qwen25vl_vision_lora_files):
config = TestConfig( config = TestConfig(
model_path=QWEN25VL_MODEL_PATH, model_path=QWEN25VL_MODEL_PATH,
-27
View File
@@ -124,30 +124,3 @@ def test_whisper_multi_lora(whisper_lora_files):
f"Expected same outputs for same adapter with different IDs. " f"Expected same outputs for same adapter with different IDs. "
f"Got: {outputs_lora1} vs {outputs_lora2}" f"Got: {outputs_lora1} vs {outputs_lora2}"
) )
@create_new_process_for_each_test()
def test_whisper_with_and_without_lora(whisper_lora_files):
"""Test that Whisper produces different outputs with and without LoRA.
This test verifies that the LoRA adapter actually affects the model output.
"""
llm = create_whisper_llm(enable_lora=True)
# Run with LoRA
outputs_with_lora = run_whisper_inference(
llm, lora_path=whisper_lora_files, lora_id=1
)
# Run without LoRA (base model only)
outputs_without_lora = run_whisper_inference(llm, lora_path=None)
# Both should produce valid outputs
assert len(outputs_with_lora[0]) > 0
assert len(outputs_without_lora[0]) > 0
print(f"Output with LoRA: {outputs_with_lora[0]}")
print(f"Output without LoRA: {outputs_without_lora[0]}")
# Note: Outputs may or may not differ depending on the adapter
# The main verification is that both configurations work