fef18b60bc
Refactor E2E CI to make it clear and faster
1. remove some uesless e2e test
2. remove some uesless function
3. Make sure all test runs with VLLMRunner to avoid oom error
4. Make sure all ops test end with torch.empty_cache to avoid oom error
5. run the test one by one to avoid resource limit error
- vLLM version: v0.10.1.1
- vLLM main:
a344a5aa0a
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
95 lines
3.5 KiB
Python
95 lines
3.5 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
import pytest
|
|
|
|
from tests.e2e.conftest import VllmRunner
|
|
from tests.e2e.model_utils import check_outputs_equal
|
|
from vllm_ascend.ascend_config import clear_ascend_config
|
|
|
|
MODEL = "Qwen/Qwen3-0.6B"
|
|
|
|
|
|
def test_concurrent_partial_prefill():
|
|
with VllmRunner(MODEL,
|
|
additional_config={
|
|
'ascend_scheduler_config': {
|
|
'enabled': True,
|
|
},
|
|
},
|
|
max_num_seqs=3,
|
|
max_num_batched_tokens=2048,
|
|
enforce_eager=True,
|
|
max_model_len=2048,
|
|
gpu_memory_utilization=0.7) as vllm_model:
|
|
outputs = vllm_model.model.generate(["Hello my name is Robert and I"] *
|
|
3)
|
|
assert len(outputs) == 3
|
|
for output in outputs:
|
|
assert len(output.outputs) == 1
|
|
|
|
clear_ascend_config()
|
|
|
|
|
|
def test_prefix_cache_stats_is_recorded():
|
|
with VllmRunner(MODEL,
|
|
additional_config={
|
|
'ascend_scheduler_config': {
|
|
'enabled': True,
|
|
},
|
|
},
|
|
max_num_seqs=3,
|
|
max_num_batched_tokens=2048,
|
|
enforce_eager=True,
|
|
max_model_len=2048,
|
|
gpu_memory_utilization=0.7) as vllm_model:
|
|
# 17 tokens will make sure first 16 tokens are cached in a block
|
|
input_tokens = {"prompt_token_ids": [101] * 129}
|
|
_ = vllm_model.model.generate([input_tokens])
|
|
outputs = vllm_model.model.generate([input_tokens])
|
|
assert outputs[0].num_cached_tokens == 128
|
|
|
|
clear_ascend_config()
|
|
|
|
|
|
@pytest.mark.parametrize("max_tokens",
|
|
[4]) # cannot align results when max_tokens > 4
|
|
@pytest.mark.parametrize("chunked_prefill_token_size", [16])
|
|
def test_chunked_prefill_with_ascend_scheduler(
|
|
max_tokens: int, chunked_prefill_token_size: int) -> None:
|
|
example_prompts = [
|
|
"vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs."
|
|
]
|
|
max_num_seqs = chunked_prefill_token_size
|
|
max_num_batched_tokens = chunked_prefill_token_size
|
|
with VllmRunner(MODEL,
|
|
additional_config={
|
|
'ascend_scheduler_config': {
|
|
'enabled': True,
|
|
'enable_chunked_prefill': True,
|
|
},
|
|
},
|
|
max_num_seqs=max_num_seqs,
|
|
max_num_batched_tokens=max_num_batched_tokens,
|
|
max_model_len=2048,
|
|
gpu_memory_utilization=0.7) as vllm_model:
|
|
chunked_prefill_output = vllm_model.generate_greedy(
|
|
example_prompts, max_tokens)
|
|
|
|
with VllmRunner(MODEL,
|
|
additional_config={
|
|
'ascend_scheduler_config': {
|
|
'enabled': True,
|
|
},
|
|
},
|
|
max_model_len=2048,
|
|
gpu_memory_utilization=0.7) as vllm_model:
|
|
vllm_output = vllm_model.generate_greedy(example_prompts, max_tokens)
|
|
|
|
check_outputs_equal(
|
|
outputs_0_lst=vllm_output,
|
|
outputs_1_lst=chunked_prefill_output,
|
|
name_0="vllm_output",
|
|
name_1="chunked_prefill_output",
|
|
)
|
|
clear_ascend_config()
|