EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[test] add w4a8 accuracy case (#5110)

Browse Source 
### What this PR does / why we need it?

This PR add w4a8  accuracy testcase for e2e test

### Does this PR introduce _any_ user-facing change?

No

### How was this patch tested?

By running the test

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

---------

Signed-off-by: cuikai (C) <c00827167@china.huawei.com>
Co-authored-by: cuikai (C) <c00827167@china.huawei.com>
This commit is contained in:
ck-hw-1018
2025-12-18 14:10:14 +08:00
committed by GitHub
parent 39fb9e7c83
commit 71e544e259
2 changed files with 34 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
.github/workflows/_e2e_test.yaml vendored
View File

@@ -192,11 +192,11 @@ jobs:
# To avoid oom, we need to run the test in a single process.
pytest -sv --durations=0 tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_DeepSeek_multistream_moe
pytest -sv --durations=0 tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_Qwen3_W4A8DYNAMIC
pytest -sv --durations=0 tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_DeepSeek_W4A8DYNAMIC
pytest -sv --durations=0 tests/e2e/multicard/test_offline_inference_distributed.py::test_sp_for_qwen3_moe
pytest -sv --durations=0 tests/e2e/multicard/test_offline_inference_distributed.py::test_fc2_for_qwen3_moe
pytest -sv --durations=0 tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_Qwen_Dense_with_flashcomm_v1
pytest -sv --durations=0 tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_Qwen_Dense_with_prefetch_mlp_weight
pytest -sv --durations=0 tests/e2e/multicard/test_offline_inference_distributed.py::test_deepseek_w4a8_accuracy
pytest -sv --durations=0 tests/e2e/multicard/test_prefix_caching.py
pytest -sv --durations=0 tests/e2e/multicard/test_pipeline_parallel.py
@@ -265,7 +265,6 @@ jobs:
VLLM_USE_MODELSCOPE: True
run: |
pytest -sv --durations=0 tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_DeepSeek_multistream_moe
pytest -sv --durations=0 tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_DeepSeek_W4A8DYNAMIC
pytest -sv --durations=0 tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_Kimi_K2_Thinking_W4A16
pytest -sv --durations=0 tests/e2e/multicard/test_data_parallel_tp2.py

49
tests/e2e/multicard/test_offline_inference_distributed.py
View File

@@ -28,6 +28,7 @@ from modelscope import snapshot_download # type: ignore
from vllm import SamplingParams
from tests.e2e.conftest import VllmRunner
from tests.e2e.model_utils import check_outputs_equal
os.environ["PYTORCH_NPU_ALLOC_CONF"] = "max_split_size_mb:256"
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
@@ -84,22 +85,6 @@ def test_models_distributed_Qwen3_W4A8DYNAMIC(model):
vllm_model.generate_greedy(prompts, max_tokens)
@pytest.mark.parametrize("model", DEEPSEEK_W4A8_MODELS)
@patch.dict(os.environ, {"HCCL_BUFFSIZE": "1024"})
def test_models_distributed_DeepSeek_W4A8DYNAMIC(model):
prompts = [
"Hello, my name is",
]
max_tokens = 5
with VllmRunner(snapshot_download(model),
dtype="auto",
tensor_parallel_size=2,
quantization="ascend",
enforce_eager=True,
enable_expert_parallel=True) as vllm_model:
vllm_model.generate_greedy(prompts, max_tokens)
def test_sp_for_qwen3_moe() -> None:
example_prompts = [
"Hello, my name is",
@@ -121,6 +106,38 @@ def test_sp_for_qwen3_moe() -> None:
vllm_model.generate(example_prompts, sampling_params)
@pytest.mark.parametrize("model", DEEPSEEK_W4A8_MODELS)
@patch.dict(os.environ, {"HCCL_BUFFSIZE": "2048"})
def test_deepseek_w4a8_accuracy(model):
prompts = [
"Hello, my name is", "The president of the United States is",
"vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs"
]
vllm_ds_w4a8_answers = [
'逍遙而至地去 accrued', '平行于我udo madreHelen', 'ysteepaolis backwards Kj'
]
sampling_params = SamplingParams(max_tokens=5, temperature=0.0)
with VllmRunner(snapshot_download(model),
dtype="auto",
tensor_parallel_size=2,
quantization="ascend",
enable_expert_parallel=True) as vllm_model:
vllm_quant_outputs = vllm_model.model.generate(prompts,
sampling_params)
vllm_quant_outputs_list = []
for output in vllm_quant_outputs:
vllm_quant_outputs_list.append(
([output.outputs[0].index], output.outputs[0].text))
vllm_answer_list = []
vllm_answer_list = ([([0], answer) for answer in vllm_ds_w4a8_answers])
check_outputs_equal(outputs_0_lst=vllm_answer_list,
outputs_1_lst=vllm_quant_outputs_list,
name_0="vllm_quant_outputs",
name_1="vllm_answer_outputs")
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_FLASHCOMM1": "1"})
@patch.dict(os.environ, {"VLLM_ASCEND_FLASHCOMM2_PARALLEL_SIZE": "1"})
def test_fc2_for_qwen3_moe() -> None: