From 9de25c294b92e42a12d1fbbb3ab3f633fa80291c Mon Sep 17 00:00:00 2001 From: Jee Jee Li Date: Wed, 27 Aug 2025 13:51:50 +0800 Subject: [PATCH] [CI/Build] Remove redundant LoRA model tests (#23706) Signed-off-by: Jee Jee Li --- tests/lora/conftest.py | 5 -- tests/lora/test_baichuan.py | 112 ------------------------------------ tests/lora/test_phi.py | 71 ----------------------- 3 files changed, 188 deletions(-) delete mode 100644 tests/lora/test_baichuan.py delete mode 100644 tests/lora/test_phi.py diff --git a/tests/lora/conftest.py b/tests/lora/conftest.py index cba573b63c045..3475993ff8f07 100644 --- a/tests/lora/conftest.py +++ b/tests/lora/conftest.py @@ -216,11 +216,6 @@ def tinyllama_lora_files(): return snapshot_download(repo_id="jashing/tinyllama-colorist-lora") -@pytest.fixture(scope="session") -def phi2_lora_files(): - return snapshot_download(repo_id="isotr0py/phi-2-test-sql-lora") - - @pytest.fixture def reset_default_device(): """ diff --git a/tests/lora/test_baichuan.py b/tests/lora/test_baichuan.py deleted file mode 100644 index 774ebb9db2106..0000000000000 --- a/tests/lora/test_baichuan.py +++ /dev/null @@ -1,112 +0,0 @@ -# SPDX-License-Identifier: Apache-2.0 -# SPDX-FileCopyrightText: Copyright contributors to the vLLM project - -import pytest - -import vllm -from vllm.distributed import cleanup_dist_env_and_memory -from vllm.lora.request import LoRARequest - -MODEL_PATH = "baichuan-inc/Baichuan-7B" - -PROMPT_TEMPLATE = """I want you to act as a SQL terminal in front of an example database, you need only to return the sql command to me.Below is an instruction that describes a task, Write a response that appropriately completes the request.\n"\n##Instruction:\nconcert_singer contains tables such as stadium, singer, concert, singer_in_concert. Table stadium has columns such as Stadium_ID, Location, Name, Capacity, Highest, Lowest, Average. Stadium_ID is the primary key.\nTable singer has columns such as Singer_ID, Name, Country, Song_Name, Song_release_year, Age, Is_male. Singer_ID is the primary key.\nTable concert has columns such as concert_ID, concert_Name, Theme, Stadium_ID, Year. concert_ID is the primary key.\nTable singer_in_concert has columns such as concert_ID, Singer_ID. concert_ID is the primary key.\nThe Stadium_ID of concert is the foreign key of Stadium_ID of stadium.\nThe Singer_ID of singer_in_concert is the foreign key of Singer_ID of singer.\nThe concert_ID of singer_in_concert is the foreign key of concert_ID of concert.\n\n###Input:\n{query}\n\n###Response:""" # noqa: E501 - - -def do_sample(llm: vllm.LLM, lora_path: str, lora_id: int) -> list[str]: - prompts = [ - PROMPT_TEMPLATE.format(query="How many singers do we have?"), - PROMPT_TEMPLATE.format( - query= - "What is the average, minimum, and maximum age of all singers from France?" # noqa: E501 - ), - PROMPT_TEMPLATE.format( - query= - "Show name, country, age for all singers ordered by age from the oldest to the youngest." # noqa: E501 - ), - ] - print(prompts) - sampling_params = vllm.SamplingParams(temperature=0, max_tokens=256) - outputs = llm.generate( - prompts, - sampling_params, - lora_request=LoRARequest(str(lora_id), lora_id, lora_path) - if lora_id else None) - # Print the outputs. - generated_texts: list[str] = [] - for output in outputs: - prompt = output.prompt - generated_text = output.outputs[0].text.strip() - generated_texts.append(generated_text) - print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") - return generated_texts - - -def test_baichuan_lora(baichuan_lora_files): - llm = vllm.LLM(MODEL_PATH, - max_model_len=1024, - enable_lora=True, - max_loras=4, - max_lora_rank=64, - trust_remote_code=True) - - expected_lora_output = [ - "SELECT count(*) FROM singer", - "SELECT avg(age) , min(age) , max(age) FROM singer WHERE Country = 'France'", # noqa: E501 - "SELECT name , country , age FROM singer ORDER BY age ASC", - ] - - output1 = do_sample(llm, baichuan_lora_files, lora_id=1) - for i in range(len(expected_lora_output)): - assert output1[i] == expected_lora_output[i] - output2 = do_sample(llm, baichuan_lora_files, lora_id=2) - for i in range(len(expected_lora_output)): - assert output2[i] == expected_lora_output[i] - - -@pytest.mark.parametrize("fully_sharded", [True, False]) -def test_baichuan_tensor_parallel_equality(baichuan_lora_files, - num_gpus_available, fully_sharded): - if num_gpus_available < 4: - pytest.skip(f"Not enough GPUs for tensor parallelism {4}") - - llm_tp1 = vllm.LLM(MODEL_PATH, - enable_lora=True, - max_num_seqs=16, - max_loras=4, - max_lora_rank=64, - trust_remote_code=True, - fully_sharded_loras=fully_sharded) - output_tp1 = do_sample(llm_tp1, baichuan_lora_files, lora_id=1) - - del llm_tp1 - cleanup_dist_env_and_memory() - - llm_tp2 = vllm.LLM(MODEL_PATH, - enable_lora=True, - max_num_seqs=16, - max_loras=4, - max_lora_rank=64, - tensor_parallel_size=2, - trust_remote_code=True, - fully_sharded_loras=fully_sharded) - output_tp2 = do_sample(llm_tp2, baichuan_lora_files, lora_id=2) - - del llm_tp2 - cleanup_dist_env_and_memory() - - assert output_tp1 == output_tp2 - - llm_tp4 = vllm.LLM(MODEL_PATH, - enable_lora=True, - max_num_seqs=16, - max_loras=4, - max_lora_rank=64, - tensor_parallel_size=4, - trust_remote_code=True, - fully_sharded_loras=fully_sharded) - output_tp4 = do_sample(llm_tp4, baichuan_lora_files, lora_id=2) - - del llm_tp4 - cleanup_dist_env_and_memory() - - assert output_tp1 == output_tp4 diff --git a/tests/lora/test_phi.py b/tests/lora/test_phi.py deleted file mode 100644 index 3090941e63679..0000000000000 --- a/tests/lora/test_phi.py +++ /dev/null @@ -1,71 +0,0 @@ -# SPDX-License-Identifier: Apache-2.0 -# SPDX-FileCopyrightText: Copyright contributors to the vLLM project - -import vllm -from vllm.lora.request import LoRARequest - -MODEL_PATH = "microsoft/phi-2" - -PROMPT_TEMPLATE = "### Instruct: {sql_prompt}\n\n### Context: {context}\n\n### Output:" # noqa: E501 - - -def do_sample(llm: vllm.LLM, lora_path: str, lora_id: int) -> list[str]: - prompts = [ - PROMPT_TEMPLATE.format( - sql_prompt= - "Which catalog publisher has published the most catalogs?", - context="CREATE TABLE catalogs (catalog_publisher VARCHAR);"), - PROMPT_TEMPLATE.format( - sql_prompt= - "Which trip started from the station with the largest dock count? Give me the trip id.", # noqa: E501 - context= - "CREATE TABLE trip (id VARCHAR, start_station_id VARCHAR); CREATE TABLE station (id VARCHAR, dock_count VARCHAR);" # noqa: E501 - ), - PROMPT_TEMPLATE.format( - sql_prompt= - "How many marine species are found in the Southern Ocean?", # noqa: E501 - context= - "CREATE TABLE marine_species (name VARCHAR(50), common_name VARCHAR(50), location VARCHAR(50));" # noqa: E501 - ), - ] - sampling_params = vllm.SamplingParams(temperature=0, - max_tokens=64, - stop="### End") - outputs = llm.generate( - prompts, - sampling_params, - lora_request=LoRARequest(str(lora_id), lora_id, lora_path) - if lora_id else None, - ) - # Print the outputs. - generated_texts: list[str] = [] - for output in outputs: - prompt = output.prompt - generated_text = output.outputs[0].text.strip() - generated_texts.append(generated_text) - print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") - return generated_texts - - -def test_phi2_lora(phi2_lora_files): - # We enable enforce_eager=True here to reduce VRAM usage for lora-test CI, - # Otherwise, the lora-test will fail due to CUDA OOM. - llm = vllm.LLM(MODEL_PATH, - max_model_len=1024, - enable_lora=True, - max_loras=2, - enforce_eager=True, - enable_chunked_prefill=True) - - expected_lora_output = [ - "SELECT catalog_publisher, COUNT(*) as num_catalogs FROM catalogs GROUP BY catalog_publisher ORDER BY num_catalogs DESC LIMIT 1;", # noqa: E501 - "SELECT trip.id FROM trip JOIN station ON trip.start_station_id = station.id WHERE station.dock_count = (SELECT MAX(dock_count) FROM station);", # noqa: E501 - "SELECT COUNT(*) FROM marine_species WHERE location = 'Southern Ocean';", # noqa: E501 - ] - - output1 = do_sample(llm, phi2_lora_files, lora_id=1) - for i in range(len(expected_lora_output)): - assert output1[i].startswith(expected_lora_output[i]) - output2 = do_sample(llm, phi2_lora_files, lora_id=2) - for i in range(len(expected_lora_output)): - assert output2[i].startswith(expected_lora_output[i])