[LoRA] Cleanup LoRA unused code (#29611)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
This commit is contained in:
Jee Jee Li
2025-11-28 22:52:58 -08:00
committed by GitHub
co-authored by Cyrus Leung
parent 4a80ad0a25
commit 39e63dec7c
46 changed files with 126 additions and 173 deletions
+8 -19
View File
@@ -154,23 +154,6 @@ def dummy_model_gate_up() -> nn.Module:
return model
@pytest.fixture(scope="session")
def llama_2_7b_base_huggingface_id():
# used as a base model for testing with sql lora adapter
return "meta-llama/Llama-2-7b-hf"
@pytest.fixture(scope="session")
def sql_lora_huggingface_id():
# huggingface repo id is used to test lora runtime downloading.
return "yard1/llama-2-7b-sql-lora-test"
@pytest.fixture(scope="session")
def sql_lora_files(sql_lora_huggingface_id):
return snapshot_download(repo_id=sql_lora_huggingface_id)
@pytest.fixture(scope="session")
def mixtral_lora_files():
# Note: this module has incorrect adapter_config.json to test
@@ -256,8 +239,14 @@ def qwen3_lora_files():
@pytest.fixture(scope="session")
def llama32_lora_files():
return snapshot_download(repo_id="jeeejeee/llama32-3b-text2sql-spider")
def llama32_lora_huggingface_id():
# huggingface repo id is used to test lora runtime downloading.
return "jeeejeee/llama32-3b-text2sql-spider"
@pytest.fixture(scope="session")
def llama32_lora_files(llama32_lora_huggingface_id):
return snapshot_download(repo_id=llama32_lora_huggingface_id)
@pytest.fixture
+7 -14
View File
@@ -26,8 +26,7 @@ def test_load_checkpoints(
chatglm3_lora_files,
):
packed_modules_mapping = BaiChuanBaseForCausalLM.packed_modules_mapping
embedding_modules = BaiChuanBaseForCausalLM.embedding_modules
embed_padding_modules = BaiChuanBaseForCausalLM.embedding_padding_modules
expected_lora_lst: list[str] = []
for module in BAICHUAN_LORA_MODULES:
if module in packed_modules_mapping:
@@ -47,8 +46,7 @@ def test_load_checkpoints(
peft_helper=peft_helper,
lora_model_id=1,
device="cpu",
embedding_modules=embedding_modules,
embedding_padding_modules=embed_padding_modules,
model_vocab_size=64000,
)
elif lora_name == "baichuan7B-zero":
# Test that the target_modules contain prefix
@@ -63,8 +61,7 @@ def test_load_checkpoints(
peft_helper=peft_helper,
lora_model_id=1,
device="cpu",
embedding_modules=embedding_modules,
embedding_padding_modules=embed_padding_modules,
model_vocab_size=64000,
)
elif lora_name == "baichuan7B-zero-regex":
# Test that the `target_modules` in the form of regular expressions,
@@ -78,8 +75,7 @@ def test_load_checkpoints(
peft_helper=peft_helper,
lora_model_id=1,
device="cpu",
embedding_modules=embedding_modules,
embedding_padding_modules=embed_padding_modules,
model_vocab_size=64000,
)
else:
# For the baichuan7B model, load chatglm3-6b's LoRA,
@@ -95,15 +91,13 @@ def test_load_checkpoints(
peft_helper=peft_helper,
lora_model_id=1,
device="cpu",
embedding_modules=embedding_modules,
embedding_padding_modules=embed_padding_modules,
model_vocab_size=64000,
)
def test_lora_weights_mapping(baichuan_lora_files):
packed_modules_mapping = BaiChuanBaseForCausalLM.packed_modules_mapping
embedding_modules = BaiChuanBaseForCausalLM.embedding_modules
embed_padding_modules = BaiChuanBaseForCausalLM.embedding_padding_modules
expected_lora_lst: list[str] = []
for module in BAICHUAN_LORA_MODULES:
if module in packed_modules_mapping:
@@ -128,8 +122,7 @@ def test_lora_weights_mapping(baichuan_lora_files):
peft_helper=peft_helper,
lora_model_id=1,
device="cpu",
embedding_modules=embedding_modules,
embedding_padding_modules=embed_padding_modules,
model_vocab_size=64000,
weights_mapper=hf_to_vllm_mapper,
)
for name in lora_model.loras:
+4 -7
View File
@@ -6,10 +6,10 @@ import pytest
from vllm.lora.models import LoRAModel
from vllm.lora.peft_helper import PEFTHelper
from vllm.lora.utils import get_adapter_absolute_path
from vllm.model_executor.models.llama import LlamaForCausalLM
from vllm.model_executor.models.qwen3 import Qwen3ForCausalLM
# Provide absolute path and huggingface lora ids
lora_fixture_name = ["sql_lora_files", "sql_lora_huggingface_id"]
lora_fixture_name = ["llama32_lora_files", "llama32_lora_huggingface_id"]
LLAMA_LORA_MODULES = [
"qkv_proj",
"o_proj",
@@ -23,9 +23,8 @@ LLAMA_LORA_MODULES = [
@pytest.mark.parametrize("lora_fixture_name", lora_fixture_name)
def test_load_checkpoints_from_huggingface(lora_fixture_name, request):
lora_name = request.getfixturevalue(lora_fixture_name)
packed_modules_mapping = LlamaForCausalLM.packed_modules_mapping
embedding_modules = LlamaForCausalLM.embedding_modules
embed_padding_modules = LlamaForCausalLM.embedding_padding_modules
packed_modules_mapping = Qwen3ForCausalLM.packed_modules_mapping
expected_lora_lst: list[str] = []
for module in LLAMA_LORA_MODULES:
if module in packed_modules_mapping:
@@ -43,8 +42,6 @@ def test_load_checkpoints_from_huggingface(lora_fixture_name, request):
peft_helper=peft_helper,
lora_model_id=1,
device="cpu",
embedding_modules=embedding_modules,
embedding_padding_modules=embed_padding_modules,
)
# Assertions to ensure the model is loaded correctly
+6 -11
View File
@@ -34,7 +34,6 @@ EMBEDDING_MODULES = {
"lm_head": "output_embeddings",
}
EMBEDDING_PADDING_MODULES = ["lm_head"]
DEVICES = (
[f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
@@ -46,24 +45,22 @@ DEFAULT_DTYPE = torch.get_default_dtype()
@pytest.mark.parametrize("device", DEVICES)
def test_from_lora_tensors(sql_lora_files, device):
tensors = load_file(os.path.join(sql_lora_files, "adapter_model.safetensors"))
def test_from_lora_tensors(qwen3_lora_files, device):
tensors = load_file(os.path.join(qwen3_lora_files, "adapter_model.safetensors"))
peft_helper = PEFTHelper.from_local_dir(
sql_lora_files, max_position_embeddings=4096
qwen3_lora_files, max_position_embeddings=4096
)
lora_model = LoRAModel.from_lora_tensors(
1,
tensors,
peft_helper=peft_helper,
device=device,
embedding_modules=EMBEDDING_MODULES,
embedding_padding_modules=EMBEDDING_PADDING_MODULES,
)
for module_name, lora in lora_model.loras.items():
assert lora.module_name == module_name
assert lora.rank == 8
assert lora.lora_alpha == 16
assert lora.lora_alpha == 32
assert lora.lora_a is not None
assert lora.lora_b is not None
assert lora.lora_a.device == torch.device(device)
@@ -430,7 +427,7 @@ def test_lru_cache_worker_adapter_manager(dist_init, dummy_model, device, tmp_pa
vllm_config.scheduler_config.max_num_seqs = 4
vllm_config.scheduler_config.max_num_batched_tokens = 2
worker_adapter_manager = LRUCacheWorkerLoRAManager(
vllm_config, device, EMBEDDING_MODULES, EMBEDDING_PADDING_MODULES
vllm_config, device, EMBEDDING_MODULES
)
worker_adapter_manager.max_num_seqs = 4
@@ -533,9 +530,7 @@ def test_worker_adapter_manager(dist_init, dummy_model_gate_up, device, tmp_path
vllm_config.scheduler_config.max_num_seqs = 4
vllm_config.scheduler_config.max_num_batched_tokens = 2
worker_adapter_manager = WorkerLoRAManager(
vllm_config, device, EMBEDDING_MODULES, EMBEDDING_PADDING_MODULES
)
worker_adapter_manager = WorkerLoRAManager(vllm_config, device, EMBEDDING_MODULES)
worker_adapter_manager.vocab_size = dummy_model_gate_up.unpadded_vocab_size
worker_adapter_manager.create_lora_manager(dummy_model_gate_up)
+17 -5
View File
@@ -40,7 +40,10 @@ EXPECTED_BASE_MODEL_OUTPUT = [
def generate_and_test(
llm: vllm.LLM, lora_path: str, lora_id: list[int | None] | int | None
llm: vllm.LLM,
lora_path: str,
lora_id: list[int | None] | int | None,
compare_lower: bool = False,
) -> None:
prompts = [
PROMPT_TEMPLATE.format(context="How many candidates are there?"),
@@ -74,12 +77,18 @@ def generate_and_test(
for i in range(len(EXPECTED_LORA_OUTPUT)):
req_lora_id = lora_id[i] if isinstance(lora_id, list) else lora_id
generated_text = generated_texts[i]
expected_output = (
EXPECTED_LORA_OUTPUT[i]
if req_lora_id is not None
else EXPECTED_BASE_MODEL_OUTPUT[i]
)
assert generated_texts[i].startswith(expected_output)
if compare_lower:
generated_text = generated_text.lower()
expected_output = expected_output.lower()
assert generated_text.startswith(expected_output)
def test_olmoe_lora(olmoe_lora_files):
@@ -146,6 +155,9 @@ def test_olmoe_lora_tp4(olmoe_lora_files, fully_sharded_loras):
tensor_parallel_size=4,
fully_sharded_loras=fully_sharded_loras,
)
generate_and_test(llm, olmoe_lora_files, lora_id=1)
generate_and_test(llm, olmoe_lora_files, lora_id=2)
generate_and_test(
llm, olmoe_lora_files, lora_id=1, compare_lower=fully_sharded_loras
)
generate_and_test(
llm, olmoe_lora_files, lora_id=2, compare_lower=fully_sharded_loras
)
+15 -13
View File
@@ -25,31 +25,33 @@ ERROR_CASES = [
]
def test_peft_helper_pass(sql_lora_files, tmp_path):
def test_peft_helper_pass(llama32_lora_files, tmp_path):
peft_helper = PEFTHelper.from_local_dir(
sql_lora_files, max_position_embeddings=4096
llama32_lora_files, max_position_embeddings=4096
)
lora_config = LoRAConfig(max_lora_rank=16, max_cpu_loras=3, max_loras=2)
peft_helper.validate_legal(lora_config)
assert peft_helper.r == 8
assert peft_helper.lora_alpha == 16
assert peft_helper.target_modules == [
"q_proj",
"v_proj",
"k_proj",
"o_proj",
"gate_proj",
"up_proj",
assert peft_helper.lora_alpha == 32
target_modules = sorted(peft_helper.target_modules)
assert target_modules == [
"down_proj",
"embed_tokens",
"gate_proj",
"k_proj",
"lm_head",
"o_proj",
"q_proj",
"up_proj",
"v_proj",
]
assert peft_helper.vllm_max_position_embeddings == 4096
# test RSLoRA
rslora_config = dict(use_rslora=True)
test_dir = tmp_path / "test_rslora"
shutil.copytree(sql_lora_files, test_dir)
shutil.copytree(llama32_lora_files, test_dir)
# Load and modify configuration
config_path = test_dir / "adapter_config.json"
@@ -70,14 +72,14 @@ def test_peft_helper_pass(sql_lora_files, tmp_path):
@pytest.mark.parametrize("test_name,config_change,expected_error", ERROR_CASES)
def test_peft_helper_error(
sql_lora_files,
llama32_lora_files,
tmp_path,
test_name: str,
config_change: dict,
expected_error: str,
):
test_dir = tmp_path / test_name
shutil.copytree(sql_lora_files, test_dir)
shutil.copytree(llama32_lora_files, test_dir)
# Load and modify configuration
config_path = test_dir / "adapter_config.json"