bfafe30953
### What this PR does / why we need it?
This PR updates the CI configuration and adjusts a set of end-to-end
(e2e) tests under tests/e2e/multicard, in order to refactor the test
suite and ensure compatibility with current codebase and CI workflows.
1. tests/e2e/multicard/test_prefix_caching.py: change model to Qwen3-8B
and rename the test case
2. tests/e2e/multicard/test_quantization.py: rename the test case
3. tests/e2e/multicard/test_qwen3_moe.py: remove duplicate test and
rename test cases
4. tests/e2e/multicard/test_qwen3_next.py: rename test cases and change
the W8A8 pruning model to the W8A8 model and remove the eager parameter
5. tests/e2e/multicard/test_shared_expert_dp.py: rename test case and
remove the eager parameter
6. tests/e2e/multicard/test_single_request_aclgraph.py: rename test case
and change Qwen3-30B to Qwen3-0.6B
7. tests/e2e/multicard/test_torchair_graph_mode.py: delete test cases
about torchair
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: hfadzxy <starmoon_zhang@163.com>
94 lines
3.0 KiB
Python
94 lines
3.0 KiB
Python
import os
|
|
|
|
import pytest
|
|
from vllm import SamplingParams
|
|
|
|
from tests.e2e.conftest import VllmRunner
|
|
from tests.e2e.model_utils import check_outputs_equal
|
|
|
|
MODELS = [
|
|
"deepseek-ai/DeepSeek-V2-Lite",
|
|
]
|
|
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
|
|
|
|
|
|
@pytest.mark.parametrize("model", MODELS)
|
|
def test_deepseek_v2_lite_enable_shared_expert_dp_tp2(model: str) -> None:
|
|
|
|
if 'HCCL_OP_EXPANSION_MODE' in os.environ:
|
|
del os.environ['HCCL_OP_EXPANSION_MODE']
|
|
|
|
prompts = [
|
|
"Hello, my name is", "The capital of the United States is",
|
|
"The capital of France is", "The future of AI is"
|
|
]
|
|
sampling_params = SamplingParams(max_tokens=32, temperature=0.0)
|
|
|
|
with VllmRunner(
|
|
model,
|
|
max_model_len=1024,
|
|
enforce_eager=True,
|
|
tensor_parallel_size=2,
|
|
enable_expert_parallel=True,
|
|
) as runner:
|
|
vllm_eager_outputs = runner.model.generate(prompts, sampling_params)
|
|
|
|
os.environ["VLLM_ASCEND_ENABLE_FLASHCOMM1"] = "1"
|
|
with VllmRunner(
|
|
model,
|
|
max_model_len=1024,
|
|
enforce_eager=True,
|
|
tensor_parallel_size=2,
|
|
enable_expert_parallel=True,
|
|
additional_config={
|
|
"enable_shared_expert_dp": True,
|
|
},
|
|
) as runner:
|
|
shared_expert_dp_eager_outputs = runner.model.generate(
|
|
prompts, sampling_params)
|
|
|
|
with VllmRunner(
|
|
model,
|
|
max_model_len=1024,
|
|
tensor_parallel_size=2,
|
|
enable_expert_parallel=True,
|
|
compilation_config={
|
|
"cudagraph_capture_sizes": [1, 4, 8, 16],
|
|
"cudagraph_mode": "FULL_DECODE_ONLY",
|
|
},
|
|
additional_config={
|
|
"enable_shared_expert_dp": True,
|
|
},
|
|
) as runner:
|
|
shared_expert_dp_aclgraph_outputs = runner.model.generate(
|
|
prompts, sampling_params)
|
|
|
|
vllm_eager_outputs_list = []
|
|
for output in vllm_eager_outputs:
|
|
vllm_eager_outputs_list.append(
|
|
(output.outputs[0].index, output.outputs[0].text))
|
|
|
|
shared_expert_dp_eager_outputs_list = []
|
|
for output in shared_expert_dp_eager_outputs:
|
|
shared_expert_dp_eager_outputs_list.append(
|
|
(output.outputs[0].index, output.outputs[0].text))
|
|
|
|
shared_expert_dp_aclgraph_outputs_list = []
|
|
for output in shared_expert_dp_aclgraph_outputs:
|
|
shared_expert_dp_aclgraph_outputs_list.append(
|
|
(output.outputs[0].index, output.outputs[0].text))
|
|
|
|
check_outputs_equal(
|
|
outputs_0_lst=vllm_eager_outputs_list,
|
|
outputs_1_lst=shared_expert_dp_eager_outputs_list,
|
|
name_0="vllm_eager_outputs",
|
|
name_1="shared_expert_dp_eager_outputs",
|
|
)
|
|
|
|
check_outputs_equal(
|
|
outputs_0_lst=vllm_eager_outputs_list,
|
|
outputs_1_lst=shared_expert_dp_aclgraph_outputs_list,
|
|
name_0="vllm_eager_outputs",
|
|
name_1="shared_expert_dp_aclgraph_outputs",
|
|
)
|