Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[TEST][PR5] Implementing test infra #1797

Merged
merged 8 commits into from
Feb 27, 2025
+152 −322
Merged

[TEST][PR5] Implementing test infra #1797

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
eac5ae2
init
iefode Feb 24, 2025
ab62fef
Update ov_genai_pipelines.py
iefode Feb 25, 2025
6d9faa7
fix test + restructure
iefode Feb 25, 2025
1729237
Fix import
iefode Feb 25, 2025
a5876bb
Merge branch 'pr_5' of github.com:iefode/openvino.genai into pr_5
iefode Feb 25, 2025
2d07374
Update ov_genai_pipelines.py
iefode Feb 25, 2025
1098261
Merge branch 'master' into pr_5
iefode Feb 26, 2025
0ab19cb
Merge branch 'master' into pr_5
iefode Feb 27, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
134 changes: 90 additions & 44 deletions tests/python_tests/common.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,59 +71,105 @@ def run_llm_pipeline(
ov_config=properties)


def run_llm_pipeline_with_ref(model_id: str,
prompts: List[str],
generation_config: GenerationConfig | dict,
tmp_path: Path | TemporaryDirectory = TemporaryDirectory(),
use_cb : bool = False,
streamer: StreamerWithResults | Callable | StreamerBase = None):
if type(generation_config) is dict:
generation_config = GenerationConfig(**generation_config)
# def run_llm_pipeline_with_ref(model_id: str,
# prompts: List[str],
# generation_config: GenerationConfig | dict,
# tmp_path: Path | TemporaryDirectory = TemporaryDirectory(),
# use_cb : bool = False,
# streamer: StreamerWithResults | Callable | StreamerBase = None):
# if type(generation_config) is dict:
# generation_config = GenerationConfig(**generation_config)

opt_model, hf_tokenizer, models_path = download_and_convert_model(model_id, Path(tmp_path.name))
# opt_model, hf_tokenizer, models_path = download_and_convert_model(model_id, Path(tmp_path.name))

ov_results = run_llm_pipeline(models_path, prompts, generation_config, use_cb, streamer=streamer.accumulate if isinstance(streamer, StreamerWithResults) else streamer)
hf_results = run_hugging_face(opt_model, hf_tokenizer, prompts, generation_config)
# ov_results = run_llm_pipeline(models_path, prompts, generation_config, use_cb, streamer=streamer.accumulate if isinstance(streamer, StreamerWithResults) else streamer)
# hf_results = run_hugging_face(opt_model, hf_tokenizer, prompts, generation_config)

compare_generation_results(prompts, hf_results, ov_results, generation_config)
# compare_generation_results(prompts, hf_results, ov_results, generation_config)


def run_cb_pipeline_with_ref(tmp_path: str,
model_id: str,
scheduler_params: dict = {},
generation_config : GenerationConfig | dict = None):
prompts, generation_configs = get_test_dataset()
scheduler_config = dict_to_scheduler_config(scheduler_params)
# def run_cb_pipeline_with_ref(tmp_path: str,
# model_id: str,
# scheduler_params: dict = {},
# generation_config : GenerationConfig | dict = None):
# prompts, generation_configs = get_test_dataset()
# scheduler_config = dict_to_scheduler_config(scheduler_params)

# override dataset's generation config
if generation_config is not None:
if type(generation_config) is dict:
generation_config = GenerationConfig(**generation_config)
generation_configs = [generation_config] * len(prompts)
# # override dataset's generation config
# if generation_config is not None:
# if type(generation_config) is dict:
# generation_config = GenerationConfig(**generation_config)
# generation_configs = [generation_config] * len(prompts)

opt_model, hf_tokenizer, models_path = download_and_convert_model(model_id, tmp_path)
# opt_model, hf_tokenizer, models_path = download_and_convert_model(model_id, tmp_path)

hf_results = run_hugging_face(opt_model, hf_tokenizer, prompts, generation_configs)
ov_results = run_continuous_batching(models_path, scheduler_config, prompts, generation_configs)
# hf_results = run_hugging_face(opt_model, hf_tokenizer, prompts, generation_configs)
# ov_results = run_continuous_batching(models_path, scheduler_config, prompts, generation_configs)

compare_generation_results(prompts, hf_results, ov_results, generation_configs)
# compare_generation_results(prompts, hf_results, ov_results, generation_configs)


# TODO: remove after Generator property is supported by LLMPipeline / VLMPipeline
def generate_and_compare_with_reference_text(models_path: Path,
prompts: List[str],
reference_texts_per_prompt: List[List[str]],
generation_configs: List[GenerationConfig],
scheduler_config: SchedulerConfig):
ov_results : List[GenerationResult] = run_continuous_batching(models_path, scheduler_config, prompts, generation_configs)

assert len(prompts) == len(reference_texts_per_prompt)
assert len(prompts) == len(ov_results)

for prompt, ref_texts_for_this_prompt, ov_result in zip(prompts, reference_texts_per_prompt, ov_results):
print(f"Prompt = {prompt}\nref text = {ref_texts_for_this_prompt}\nOV result = {ov_result.m_generation_ids}")

assert len(ref_texts_for_this_prompt) == len(ov_result.m_generation_ids)
for ref_text, ov_text in zip(ref_texts_for_this_prompt, ov_result.m_generation_ids):
assert ref_text == ov_text

# def generate_and_compare_with_reference_text(models_path: Path,
# prompts: List[str],
# reference_texts_per_prompt: List[List[str]],
# generation_configs: List[GenerationConfig],
# scheduler_config: SchedulerConfig):
# ov_results : List[GenerationResult] = run_continuous_batching(models_path, scheduler_config, prompts, generation_configs)

# assert len(prompts) == len(reference_texts_per_prompt)
# assert len(prompts) == len(ov_results)

# for prompt, ref_texts_for_this_prompt, ov_result in zip(prompts, reference_texts_per_prompt, ov_results):
# print(f"Prompt = {prompt}\nref text = {ref_texts_for_this_prompt}\nOV result = {ov_result.m_generation_ids}")

# assert len(ref_texts_for_this_prompt) == len(ov_result.m_generation_ids)
# for ref_text, ov_text in zip(ref_texts_for_this_prompt, ov_result.m_generation_ids):
# assert ref_text == ov_text


def generate_and_compare(model: Path | str,
prompts : List[str],
generation_config: List[GenerationConfig] | GenerationConfig | dict,
pipeline_type: PipelineType = PipelineType.PAGED_ATTENTION,
scheduler_config: SchedulerConfig | dict = SchedulerConfig(),
ref : List[List[str]] = None,
streamer: StreamerWithResults | Callable | StreamerBase = None,
tmp_path: Path | TemporaryDirectory = TemporaryDirectory()) :
if type(generation_config) is dict:
gen_config = GenerationConfig(**generation_config)
elif isinstance(generation_config, GenerationConfig):
gen_config = [generation_config] * len(prompts)
else:
gen_config = generation_config

if isinstance(scheduler_config, SchedulerConfig):
scheduler_config_ = scheduler_config
else:
scheduler_config_= dict_to_scheduler_config(scheduler_config)

if isinstance(model, Path):
models_path = model
else:
opt_model, hf_tokenizer, models_path = download_and_convert_model(model, Path(tmp_path.name))

ov_results = run_ov_pipeline(models_path=models_path,
prompt=prompts,
generation_config=gen_config,
pipeline_type=pipeline_type,
streamer=streamer.accumulate if isinstance(streamer, StreamerWithResults) else streamer,
scheduler_config=scheduler_config_,
ov_config=get_default_llm_properties())
if ref is None:
ref_results = run_hugging_face(opt_model, hf_tokenizer, prompts, generation_config)
compare_generation_results(prompts, ref_results, ov_results, generation_config)
else:
assert len(prompts) == len(ref)
assert len(prompts) == len(ov_results)

for prompt, ref_texts_for_this_prompt, ov_result in zip(prompts, ref, ov_results):
print(f"Prompt = {prompt}\nref text = {ref_texts_for_this_prompt}\nOV result = {ov_result.m_generation_ids}")

assert len(ref_texts_for_this_prompt) == len(ov_result.m_generation_ids)
for ref_text, ov_text in zip(ref_texts_for_this_prompt, ov_result.m_generation_ids):
assert ref_text == ov_text
160 changes: 0 additions & 160 deletions tests/python_tests/ov_genai_test_utils.py
View file
Open in desktop

This file was deleted.

27 changes: 20 additions & 7 deletions tests/python_tests/test_continuous_batching.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@

from openvino_genai import ContinuousBatchingPipeline, LLMPipeline, GenerationConfig, SchedulerConfig, draft_model

from common import generate_and_compare_with_reference_text, run_cb_pipeline_with_ref
from common import generate_and_compare
from test_sampling import RandomSamplingTestStruct, get_current_platform_ref_texts

from utils.generation_config import get_greedy, get_beam_search, \
Expand Down Expand Up @@ -40,19 +40,19 @@ def read_models_list(file_name: str):
@pytest.mark.precommit
@pytest.mark.parametrize("model_id", read_models_list(os.path.join(os.path.dirname(os.path.realpath(__file__)), "models", "precommit")))
def test_e2e_precommit(tmp_path, model_id):
run_cb_pipeline_with_ref(tmp_path, model_id)
generate_and_compare(tmp_path=tmp_path, model=model_id)


@pytest.mark.nightly
@pytest.mark.parametrize("model_id", read_models_list(os.path.join(os.path.dirname(os.path.realpath(__file__)), "models", "nightly")))
def test_e2e_nightly(tmp_path, model_id):
run_cb_pipeline_with_ref(tmp_path, model_id)
generate_and_compare(tmp_path=tmp_path, model=model_id)


@pytest.mark.real_models
@pytest.mark.parametrize("model_id", read_models_list(os.path.join(os.path.dirname(os.path.realpath(__file__)), "models", "real_models")))
def test_e2e_real_models(tmp_path, model_id):
run_cb_pipeline_with_ref(tmp_path, model_id)
generate_and_compare(tmp_path=tmp_path, model=model_id)

#
# Comparison with stateful
Expand Down Expand Up @@ -208,8 +208,11 @@ def get_beam_search_seq_len_300() -> GenerationConfig:
@pytest.mark.parametrize("params", scheduler_params_list)
@pytest.mark.precommit
def test_preemption(tmp_path, params):
run_cb_pipeline_with_ref(tmp_path, "facebook/opt-125m", scheduler_params=params[0], generation_config=params[1])
model_id = "facebook/opt-125m"
scheduler_params = params[0]
generation_config = params[1]

generate_and_compare(tmp_path=tmp_path, model=model_id, scheduler_config=scheduler_params, generation_config=generation_config)

multinomial_params = RandomSamplingTestStruct(
generation_config=[
Expand Down Expand Up @@ -261,7 +264,12 @@ def test_preemption_with_multinomial(tmp_path, dynamic_split_fuse):
model, hf_tokenizer, models_path = download_and_convert_model(model_id, tmp_path)

scheduler_config = dict_to_scheduler_config({"num_kv_blocks": 3, "dynamic_split_fuse": dynamic_split_fuse, "max_num_batched_tokens": 256, "max_num_seqs": 256})
generate_and_compare_with_reference_text(models_path, multinomial_params.prompts, multinomial_params.ref_texts, generation_configs, scheduler_config)
generate_and_compare(model=models_path,
pipeline_type=PipelineType.CONTINIOUS_BATCHING,
prompts=multinomial_params.prompts,
ref=multinomial_params.ref_texts,
generation_config=generation_configs,
scheduler_config=scheduler_config)


multinomial_params_n_seq = RandomSamplingTestStruct(
Expand Down Expand Up @@ -337,7 +345,12 @@ def test_preemption_with_multinomial_n_seq(tmp_path, dynamic_split_fuse):

# needed kv_blocks - 16 (2 blocks per sequence (30 tokens to generated text + prompt (> 2 tokens)) * (1 + 3 + 4) seq )
scheduler_config = dict_to_scheduler_config({"num_kv_blocks": 8, "dynamic_split_fuse": dynamic_split_fuse, "max_num_batched_tokens": 256, "max_num_seqs": 256})
generate_and_compare_with_reference_text(models_path, multinomial_params_n_seq.prompts, multinomial_params_n_seq.ref_texts, multinomial_params_n_seq.generation_config, scheduler_config)
generate_and_compare(model=models_path,
pipeline_type=PipelineType.CONTINIOUS_BATCHING,
prompts=multinomial_params_n_seq.prompts,
ref=multinomial_params_n_seq.ref_texts,
generation_config=multinomial_params_n_seq.generation_config,
scheduler_config=scheduler_config)

def get_data_by_pipeline_type(model_path: Path, pipeline_type: str, generation_config: GenerationConfig):
device = "CPU"
Expand Down
Loading
Loading