[Core] Remove tokenizer group in vLLM (#24078)

Signed-off-by: Zhuohan Li <zhuohan123@gmail.com>
This commit is contained in:
Zhuohan Li
2025-09-17 08:42:59 +00:00
committed by GitHub
parent c15309a730
commit 6c47f6bfa4
49 changed files with 276 additions and 934 deletions
+4 -24
View File
@@ -13,14 +13,6 @@ from ..utils import VLLM_PATH, create_new_process_for_each_test, multi_gpu_test
MODEL_PATH = "meta-llama/Llama-2-7b-hf"
EXPECTED_NO_LORA_OUTPUT = [
"\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_75 (icao VARCHAR, airport VARCHAR)\n\n question: Name the ICAO for lilongwe international airport [/user] [assistant]\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_76 (icao VARCHAR, airport VARCHAR)\n\n question: Name the ICAO for lilongwe international airport [/user] [assistant]\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_77 (icao VARCHAR, airport VARCHAR)\n\n question: Name the ICAO for lilongwe international airport [/user] [assistant]\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_78 (icao VARCHAR, airport VARCHAR)\n\n question: Name the ICAO for lilongwe international airport [/user]", # noqa: E501
" Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_11 (nationality VARCHAR, elector VARCHAR)\n\n question: When Anchero Pantaleone was the elector what is under nationality? ", # noqa: E501
"\n\n answer: 1\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_96 (one_mora VARCHAR, gloss VARCHAR, accented_mora VARCHAR)\n\n question: What is the one mora for a high tone mora with a gloss of /˧kot/ [kòt]? [/user] [assistant]\n\n answer: 2\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_97 (one_mora VARCHAR, gloss VARCHAR, accented_mora VARCHAR)\n\n question: What is the one mora for a high tone mora with a gloss of /˧kot/ [kòt]? [/user] [assistant]\n\n answer: 2\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_98 (one_mora VARCHAR, gloss VARCHAR, accented_mora VARCHAR)\n\n question: What is the one m", # noqa: E501
" Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE candidate (people_id VARCHAR, unsure_rate INTEGER); CREATE TABLE people (sex VARCHAR, people_id VARCHAR)\n\n question: which gender got the highest average uncertain ratio. ", # noqa: E501
" Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_60 (pick INTEGER, former_wnba_team VARCHAR)\n\n question: What pick was a player that previously played for the Minnesota Lynx? ", # noqa: E501
"\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_28138035_4 (womens_doubles VARCHAR, mens_singles VARCHAR)\n\n question: Name the women's doubles for werner schlager [/user] [assistant]\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_28138035_4 (womens_doubles VARCHAR, mens_singles VARCHAR)\n\n question: Name the women's doubles for werner schlager [/user] [assistant]\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_28138035_4 (womens_doubles VARCHAR, mens_singles VARCHAR)\n\n question: Name the women's doubles for werner schlager [/user] [assistant]\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE", # noqa: E501
]
EXPECTED_LORA_OUTPUT = [
" SELECT icao FROM table_name_74 WHERE airport = 'lilongwe international airport' ", # noqa: E501
" SELECT nationality FROM table_name_11 WHERE elector = 'anchero pantaleone' ", # noqa: E501
@@ -79,23 +71,12 @@ def generate_and_test(llm,
sql_lora_files,
tensorizer_config_dict: Union[dict, None] = None):
print("lora adapter created")
assert do_sample(llm,
sql_lora_files,
tensorizer_config_dict=tensorizer_config_dict,
lora_id=0) == EXPECTED_NO_LORA_OUTPUT
print("lora 1")
assert do_sample(llm,
sql_lora_files,
tensorizer_config_dict=tensorizer_config_dict,
lora_id=1) == EXPECTED_LORA_OUTPUT
print("no lora")
assert do_sample(llm,
sql_lora_files,
tensorizer_config_dict=tensorizer_config_dict,
lora_id=0) == EXPECTED_NO_LORA_OUTPUT
print("lora 2")
assert do_sample(llm,
sql_lora_files,
@@ -110,6 +91,7 @@ def test_llama_lora(sql_lora_files):
llm = vllm.LLM(
MODEL_PATH,
tokenizer=sql_lora_files,
enable_lora=True,
# also test odd max_num_seqs
max_num_seqs=13,
@@ -123,6 +105,7 @@ def test_llama_lora_tp4(sql_lora_files):
llm = vllm.LLM(
MODEL_PATH,
tokenizer=sql_lora_files,
enable_lora=True,
max_num_seqs=16,
max_loras=4,
@@ -137,6 +120,7 @@ def test_llama_lora_tp4_fully_sharded_loras(sql_lora_files):
llm = vllm.LLM(
MODEL_PATH,
tokenizer=sql_lora_files,
enable_lora=True,
max_num_seqs=16,
max_loras=4,
@@ -184,6 +168,7 @@ def test_tp2_serialize_and_deserialize_lora(tmp_path, sql_lora_files,
tensorizer_config = TensorizerConfig(tensorizer_uri=str(model_uri))
loaded_llm = LLM(model=model_ref,
tokenizer=sql_lora_files,
load_format="tensorizer",
enable_lora=True,
enforce_eager=True,
@@ -195,11 +180,6 @@ def test_tp2_serialize_and_deserialize_lora(tmp_path, sql_lora_files,
tc_as_dict = tensorizer_config.to_serializable()
print("lora adapter created")
assert do_sample(loaded_llm,
sql_lora_files,
tensorizer_config_dict=tc_as_dict,
lora_id=0) == EXPECTED_NO_LORA_OUTPUT
print("lora 1")
assert do_sample(loaded_llm,
sql_lora_files,
-135
View File
@@ -1,135 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from vllm.config import CacheConfig, DeviceConfig, ModelConfig, VllmConfig
from vllm.config.lora import LoRAConfig
from vllm.lora.request import LoRARequest
from vllm.sampling_params import SamplingParams
from vllm.transformers_utils.tokenizer_group import init_tokenizer_from_configs
from vllm.v1.engine.processor import Processor
def test_allowed_token_ids_with_lora_vocab(llama_2_7b_base_huggingface_id,
sql_lora_files):
"""
Test that we properly resolve the range of allowed token ids for lora
adapters that define additional tokens.
"""
# Set up a base model compatible with the sql_lora_files adapter and
# a known number of tokens in the base model.
model_config = ModelConfig(
model=llama_2_7b_base_huggingface_id,
tokenizer=llama_2_7b_base_huggingface_id,
tokenizer_mode="auto",
)
vllm_config = VllmConfig(
model_config=model_config,
cache_config=CacheConfig(),
device_config=DeviceConfig(),
lora_config=LoRAConfig(),
)
tokenizer = init_tokenizer_from_configs(
model_config=vllm_config.model_config,
scheduler_config=vllm_config.scheduler_config,
lora_config=vllm_config.lora_config)
processor = Processor(vllm_config, tokenizer)
lora_request = LoRARequest("1", 1, str(sql_lora_files))
request_id = "1"
prompt = "a prompt"
# tokens added in the lora adapter should not raise an error
lora_token_ids = [32000, 32001, 32002, 32003]
processor.process_inputs(
request_id,
prompt,
params=SamplingParams(allowed_token_ids=lora_token_ids),
lora_request=lora_request)
# tokens in the base model should not raise an error
base_token_ids = [1000, 1001, 1002, 1003]
processor.process_inputs(
request_id,
prompt,
params=SamplingParams(allowed_token_ids=base_token_ids),
lora_request=lora_request)
# tokens not in the lora adapter should raise an error
invalid_token_ids = [35000, 35001, 35002, 35003]
with pytest.raises(ValueError):
processor.process_inputs(
request_id,
prompt,
params=SamplingParams(allowed_token_ids=invalid_token_ids),
lora_request=lora_request)
# tokens in the lora adapter with no lora request should raise an error
with pytest.raises(ValueError):
processor.process_inputs(
request_id,
prompt,
params=SamplingParams(allowed_token_ids=lora_token_ids),
)
def test_allowed_token_ids_with_lora_adapter_no_vocab(
qwen25vl_base_huggingface_id, qwen25vl_lora_files):
"""
Test that we properly resolve the range of allowed token ids for lora
adapters that do not define additional tokens.
"""
# Set up a base model compatible with the qwen25vl_lora_files adapter and
# a known number of tokens in the base model.
model_config = ModelConfig(
model=qwen25vl_base_huggingface_id,
tokenizer=qwen25vl_base_huggingface_id,
tokenizer_mode="auto",
)
vllm_config = VllmConfig(
model_config=model_config,
cache_config=CacheConfig(),
device_config=DeviceConfig(),
lora_config=LoRAConfig(),
)
tokenizer = init_tokenizer_from_configs(
model_config=vllm_config.model_config,
scheduler_config=vllm_config.scheduler_config,
lora_config=vllm_config.lora_config)
processor = Processor(vllm_config, tokenizer)
lora_request = LoRARequest("1", 1, str(qwen25vl_lora_files))
request_id = "1"
prompt = "a prompt"
# tokens in the base model should not raise an error
base_token_ids = [1000, 1001, 1002, 1003]
processor.process_inputs(
request_id,
prompt,
params=SamplingParams(allowed_token_ids=base_token_ids),
lora_request=lora_request)
# tokens in the base model with no lora request should not raise an error
base_token_ids = [1000, 1001, 1002, 1003]
processor.process_inputs(
request_id,
prompt,
params=SamplingParams(allowed_token_ids=base_token_ids),
)
# tokens not in the base model should raise an error
invalid_token_ids = [200000, 200001, 200002, 200003]
with pytest.raises(ValueError):
processor.process_inputs(
request_id,
prompt,
params=SamplingParams(allowed_token_ids=invalid_token_ids),
lora_request=lora_request)
+2 -27
View File
@@ -82,31 +82,20 @@ def test_quant_model_lora(tinyllama_lora_files, model):
gpu_memory_utilization=0.2, #avoid OOM
quantization=model.quantization,
trust_remote_code=True,
enable_chunked_prefill=True)
enable_chunked_prefill=True,
tokenizer=tinyllama_lora_files)
if model.quantization is None:
expected_no_lora_output = [
"Here are some examples of orange-brown colors",
"I'm sorry, I don't have"
]
expected_lora_output = [
"#ff8050",
"#ff8080",
]
elif model.quantization == "awq":
expected_no_lora_output = [
"I'm sorry, I don't understand",
"I'm sorry, I don't understand",
]
expected_lora_output = [
"#f07700: A v",
"#f00000: A v",
]
elif model.quantization == "gptq":
expected_no_lora_output = [
"I'm sorry, I don't have",
"I'm sorry, I don't have",
]
expected_lora_output = [
"#f08800: This is",
"#f07788 \n#",
@@ -117,7 +106,6 @@ def test_quant_model_lora(tinyllama_lora_files, model):
# Assert that the outputs changed.
if (model.quantization == "gptq"
and expected_output is expected_lora_output):
assert output != expected_no_lora_output
for i, o in enumerate(output):
assert o.startswith(
'#'), f"Expected example {i} to start with # but got {o}"
@@ -127,12 +115,6 @@ def test_quant_model_lora(tinyllama_lora_files, model):
max_tokens = 10
print("lora adapter created")
output = do_sample(llm,
tinyllama_lora_files,
lora_id=0,
max_tokens=max_tokens)
expect_match(output, expected_no_lora_output)
print("lora 1")
output = do_sample(llm,
tinyllama_lora_files,
@@ -140,13 +122,6 @@ def test_quant_model_lora(tinyllama_lora_files, model):
max_tokens=max_tokens)
expect_match(output, expected_lora_output)
print("no lora")
output = do_sample(llm,
tinyllama_lora_files,
lora_id=0,
max_tokens=max_tokens)
expect_match(output, expected_no_lora_output)
print("lora 2")
output = do_sample(llm,
tinyllama_lora_files,
-72
View File
@@ -1,72 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from transformers import AutoTokenizer, PreTrainedTokenizerBase
from vllm.lora.request import LoRARequest
from vllm.transformers_utils.tokenizer import get_lora_tokenizer
from vllm.transformers_utils.tokenizer_group import TokenizerGroup
@pytest.mark.asyncio
@pytest.mark.parametrize("tokenizer_group_type", [None, "ray"])
async def test_tokenizer_group_lora(sql_lora_files, tokenizer_group_type):
reference_tokenizer = AutoTokenizer.from_pretrained(sql_lora_files)
tokenizer_group = TokenizerGroup(
tokenizer_id="gpt2",
enable_lora=True,
max_num_seqs=1,
max_loras=1,
max_input_length=None,
)
lora_request = LoRARequest("1", 1, sql_lora_files)
assert reference_tokenizer.encode("prompt") == tokenizer_group.encode(
prompt="prompt", lora_request=lora_request)
assert reference_tokenizer.encode(
"prompt") == await tokenizer_group.encode_async(
prompt="prompt", lora_request=lora_request)
assert isinstance(tokenizer_group.get_lora_tokenizer(None),
PreTrainedTokenizerBase)
assert tokenizer_group.get_lora_tokenizer(
None) == await tokenizer_group.get_lora_tokenizer_async(None)
assert isinstance(tokenizer_group.get_lora_tokenizer(lora_request),
PreTrainedTokenizerBase)
assert tokenizer_group.get_lora_tokenizer(
lora_request) != tokenizer_group.get_lora_tokenizer(None)
assert tokenizer_group.get_lora_tokenizer(
lora_request) == await tokenizer_group.get_lora_tokenizer_async(
lora_request)
def test_get_lora_tokenizer(sql_lora_files, tmp_path):
lora_request = None
tokenizer = get_lora_tokenizer(lora_request)
assert not tokenizer
lora_request = LoRARequest("1", 1, sql_lora_files)
tokenizer = get_lora_tokenizer(lora_request)
assert tokenizer.get_added_vocab()
lora_request = LoRARequest("1", 1, str(tmp_path))
tokenizer = get_lora_tokenizer(lora_request)
assert not tokenizer
@pytest.mark.parametrize("enable_lora", [True, False])
@pytest.mark.parametrize("max_num_seqs", [1, 2])
@pytest.mark.parametrize("max_loras", [1, 2])
def test_lora_tokenizers(enable_lora, max_num_seqs, max_loras):
tokenizer_group = TokenizerGroup(
tokenizer_id="gpt2",
enable_lora=enable_lora,
max_num_seqs=max_num_seqs,
max_loras=max_loras,
max_input_length=None,
)
if enable_lora:
assert tokenizer_group.lora_tokenizers.capacity == max(
max_num_seqs, max_loras)
else:
assert tokenizer_group.lora_tokenizers.capacity == 0