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

Refactor e2e CI (#2276)

Browse Source 
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>
This commit is contained in:
wangxiyuan
2025-09-02 09:02:22 +08:00
committed by GitHub
parent 0df059f41a
commit fef18b60bc
41 changed files with 374 additions and 1757 deletions
Download Patch File Download Diff File Expand all files Collapse all files

136
tests/e2e/conftest.py
View File

@@ -33,13 +33,11 @@ from transformers import (AutoConfig, AutoModelForCausalLM, AutoTokenizer,
from transformers.models.auto.auto_factory import _BaseAutoModelClass
from vllm import LLM, SamplingParams
from vllm.config import TaskOption, _get_and_verify_dtype
from vllm.inputs import ExplicitEncoderDecoderPrompt, TextPrompt, TokensPrompt
from vllm.inputs import TextPrompt
from vllm.outputs import RequestOutput
from vllm.sampling_params import BeamSearchParams
from vllm.transformers_utils.utils import maybe_model_redirect
from vllm.utils import is_list_of
from tests.e2e.model_utils import (PROMPT_TEMPLATES, TokensTextLogprobs,
from tests.e2e.model_utils import (TokensTextLogprobs,
TokensTextLogprobsPromptLogprobs)
# TODO: remove this part after the patch merged into vllm, if
# we not explicitly patch here, some of them might be effectiveless
@@ -62,7 +60,6 @@ PromptAudioInput = _PromptMultiModalInput[Tuple[np.ndarray, int]]
PromptVideoInput = _PromptMultiModalInput[np.ndarray]
_TEST_DIR = os.path.dirname(__file__)
_TEST_PROMPTS = [os.path.join(_TEST_DIR, "prompts", "example.txt")]
def cleanup_dist_env_and_memory(shutdown_ray: bool = False):
@@ -89,13 +86,13 @@ class VllmRunner:
# Use smaller max model length, otherwise bigger model cannot run due
# to kv cache size limit.
max_model_len: int = 1024,
dtype: str = "half",
dtype: str = "auto",
disable_log_stats: bool = True,
tensor_parallel_size: int = 1,
block_size: int = 16,
enable_chunked_prefill: bool = False,
swap_space: int = 4,
enforce_eager: Optional[bool] = True,
enforce_eager: Optional[bool] = False,
quantization: Optional[str] = None,
**kwargs,
) -> None:
@@ -220,26 +217,6 @@ class VllmRunner:
if sampling_params.prompt_logprobs is None else
toks_str_logsprobs_prompt_logprobs)
def generate_encoder_decoder_w_logprobs(
self,
encoder_decoder_prompts: List[ExplicitEncoderDecoderPrompt[str, str]],
sampling_params: SamplingParams,
) -> Union[List[TokensTextLogprobs],
List[TokensTextLogprobsPromptLogprobs]]:
'''
Logprobs generation for vLLM encoder/decoder models
'''
assert sampling_params.logprobs is not None
req_outputs = self.model.generate(encoder_decoder_prompts,
sampling_params=sampling_params)
toks_str_logsprobs_prompt_logprobs = (
self._final_steps_generate_w_logprobs(req_outputs))
# Omit prompt logprobs if not required by sampling params
return ([x[0:-1] for x in toks_str_logsprobs_prompt_logprobs]
if sampling_params.prompt_logprobs is None else
toks_str_logsprobs_prompt_logprobs)
def generate_greedy(
self,
prompts: List[str],
@@ -284,53 +261,6 @@ class VllmRunner:
audios=audios,
videos=videos)
def generate_encoder_decoder_greedy_logprobs(
self,
encoder_decoder_prompts: List[ExplicitEncoderDecoderPrompt[str, str]],
max_tokens: int,
num_logprobs: int,
num_prompt_logprobs: Optional[int] = None,
) -> Union[List[TokensTextLogprobs],
List[TokensTextLogprobsPromptLogprobs]]:
greedy_logprobs_params = SamplingParams(
temperature=0.0,
max_tokens=max_tokens,
logprobs=num_logprobs,
prompt_logprobs=(num_prompt_logprobs),
)
'''
Greedy logprobs generation for vLLM encoder/decoder models
'''
return self.generate_encoder_decoder_w_logprobs(
encoder_decoder_prompts, greedy_logprobs_params)
def generate_beam_search(
self,
prompts: Union[List[str], List[List[int]]],
beam_width: int,
max_tokens: int,
) -> List[Tuple[List[List[int]], List[str]]]:
if is_list_of(prompts, str, check="all"):
prompts = [TextPrompt(prompt=prompt) for prompt in prompts]
else:
prompts = [
TokensPrompt(prompt_token_ids=tokens) for tokens in prompts
]
outputs = self.model.beam_search(
prompts,
BeamSearchParams(beam_width=beam_width, max_tokens=max_tokens))
returned_outputs = []
for output in outputs:
token_ids = [x.tokens for x in output.sequences]
texts = [x.text for x in output.sequences]
returned_outputs.append((token_ids, texts))
return returned_outputs
def classify(self, prompts: List[str]) -> List[List[float]]:
req_outputs = self.model.classify(prompts)
return [req_output.outputs.probs for req_output in req_outputs]
def encode(
self,
prompts: List[str],
@@ -346,14 +276,6 @@ class VllmRunner:
req_outputs = self.model.embed(inputs)
return [req_output.outputs.embedding for req_output in req_outputs]
def score(
self,
text_1: Union[str, List[str]],
text_2: Union[str, List[str]],
) -> List[float]:
req_outputs = self.model.score(text_1, text_2)
return [req_output.outputs.score for req_output in req_outputs]
def __enter__(self):
return self
@@ -362,35 +284,6 @@ class VllmRunner:
cleanup_dist_env_and_memory()
@pytest.fixture(scope="session")
def vllm_runner():
return VllmRunner
@pytest.fixture(params=list(PROMPT_TEMPLATES.keys()))
def prompt_template(request):
return PROMPT_TEMPLATES[request.param]
def _read_prompts(filename: str) -> list[str]:
with open(filename) as f:
prompts = f.readlines()
return prompts
@pytest.fixture
def example_prompts() -> list[str]:
prompts = []
for filename in _TEST_PROMPTS:
prompts += _read_prompts(filename)
return prompts
@pytest.fixture(scope="session")
def ilama_lora_files():
return snapshot_download(repo_id="vllm-ascend/ilama-text2sql-spider")
class HfRunner:
def get_default_device(self):
@@ -515,5 +408,22 @@ class HfRunner:
@pytest.fixture(scope="session")
def hf_runner():
return HfRunner
def ilama_lora_files():
return snapshot_download(repo_id="vllm-ascend/ilama-text2sql-spider")
def qwen_prompt(questions: List[str]) -> List[str]:
placeholder = "<|image_pad|>"
return [("<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
f"<|im_start|>user\n<|vision_start|>{placeholder}<|vision_end|>"
f"{q}<|im_end|>\n<|im_start|>assistant\n") for q in questions]
PROMPT_TEMPLATES = {
"qwen2.5vl": qwen_prompt,
}
@pytest.fixture(params=list(PROMPT_TEMPLATES.keys()))
def prompt_template(request):
return PROMPT_TEMPLATES[request.param]