[Misc] Reduce LoRA-related static variable (#13166)

This commit is contained in:
Jee Jee Li
2025-02-22 00:21:30 -08:00
committed by GitHub
parent 2cb8c1540e
commit 105b8ce4c0
41 changed files with 120 additions and 395 deletions
+15 -2
View File
@@ -23,6 +23,7 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.sampler import Sampler
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
from vllm.model_executor.model_loader import get_model
from vllm.model_executor.models.interfaces import SupportsLoRA
from vllm.platforms import current_platform
@@ -98,9 +99,13 @@ def dist_init_torch_only():
backend=backend)
class DummyLoRAModel(nn.Sequential, SupportsLoRA):
pass
@pytest.fixture
def dummy_model() -> nn.Module:
model = nn.Sequential(
model = DummyLoRAModel(
OrderedDict([
("dense1", ColumnParallelLinear(764, 100)),
("dense2", RowParallelLinear(100, 50)),
@@ -121,12 +126,13 @@ def dummy_model() -> nn.Module:
("sampler", Sampler())
]))
model.config = MagicMock()
model.embedding_modules = {"lm_head": "lm_head"}
return model
@pytest.fixture
def dummy_model_gate_up() -> nn.Module:
model = nn.Sequential(
model = DummyLoRAModel(
OrderedDict([
("dense1", ColumnParallelLinear(764, 100)),
("dense2", RowParallelLinear(100, 50)),
@@ -147,6 +153,13 @@ def dummy_model_gate_up() -> nn.Module:
("sampler", Sampler())
]))
model.config = MagicMock()
model.packed_modules_mapping = {
"gate_up_proj": [
"gate_proj",
"up_proj",
],
}
model.embedding_modules = {"lm_head": "lm_head"}
return model
+9 -4
View File
@@ -12,6 +12,12 @@ from vllm.model_executor.models.utils import WeightsMapper
lora_lst = [
"baichuan7B", "baichuan7B-zero", "baichuan7B-zero-regex", "chatglm3-6b"
]
BAICHUAN_LORA_MODULES = [
"W_pack",
"o_proj",
"gate_up_proj",
"down_proj",
]
@pytest.mark.parametrize("lora_name", lora_lst)
@@ -22,12 +28,11 @@ def test_load_checkpoints(
baichuan_regex_lora_files,
chatglm3_lora_files,
):
supported_lora_modules = BaiChuanBaseForCausalLM.supported_lora_modules
packed_modules_mapping = BaiChuanBaseForCausalLM.packed_modules_mapping
embedding_modules = BaiChuanBaseForCausalLM.embedding_modules
embed_padding_modules = BaiChuanBaseForCausalLM.embedding_padding_modules
expected_lora_modules: List[str] = []
for module in supported_lora_modules:
for module in BAICHUAN_LORA_MODULES:
if module in packed_modules_mapping:
expected_lora_modules.extend(packed_modules_mapping[module])
else:
@@ -90,12 +95,12 @@ def test_load_checkpoints(
def test_lora_weights_mapping(baichuan_lora_files):
supported_lora_modules = BaiChuanBaseForCausalLM.supported_lora_modules
packed_modules_mapping = BaiChuanBaseForCausalLM.packed_modules_mapping
embedding_modules = BaiChuanBaseForCausalLM.embedding_modules
embed_padding_modules = BaiChuanBaseForCausalLM.embedding_padding_modules
expected_lora_modules: List[str] = []
for module in supported_lora_modules:
for module in BAICHUAN_LORA_MODULES:
if module in packed_modules_mapping:
expected_lora_modules.extend(packed_modules_mapping[module])
else:
+5 -2
View File
@@ -11,17 +11,20 @@ from vllm.model_executor.models.llama import LlamaForCausalLM
# Provide absolute path and huggingface lora ids
lora_fixture_name = ["sql_lora_files", "sql_lora_huggingface_id"]
LLAMA_LORA_MODULES = [
"qkv_proj", "o_proj", "gate_up_proj", "down_proj", "embed_tokens",
"lm_head"
]
@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)
supported_lora_modules = LlamaForCausalLM.supported_lora_modules
packed_modules_mapping = LlamaForCausalLM.packed_modules_mapping
embedding_modules = LlamaForCausalLM.embedding_modules
embed_padding_modules = LlamaForCausalLM.embedding_padding_modules
expected_lora_modules: List[str] = []
for module in supported_lora_modules:
for module in LLAMA_LORA_MODULES:
if module in packed_modules_mapping:
expected_lora_modules.extend(packed_modules_mapping[module])
else:
+8 -18
View File
@@ -19,7 +19,6 @@ from vllm.lora.peft_helper import PEFTHelper
from vllm.lora.request import LoRARequest
from vllm.lora.worker_manager import (LRUCacheWorkerLoRAManager,
WorkerLoRAManager)
from vllm.model_executor.layers.linear import RowParallelLinear
from vllm.platforms import current_platform
EMBEDDING_MODULES = {
@@ -114,19 +113,16 @@ def create_packed_lora(
def test_replace_submodules(dist_init, dummy_model):
model = dummy_model
model.supported_lora_modules = ["dense1", "layer1.dense2"]
model.packed_modules_mapping = {}
manager = LoRAModelManager(
model, 1, 1, 1,
LoRAConfig(max_lora_rank=8, max_cpu_loras=8, max_loras=8),
torch.device(DEVICES[0]))
model = manager.model
assert isinstance(model.get_submodule("dense1"),
ColumnParallelLinearWithLoRA)
assert isinstance(model.get_submodule("layer1.dense1"),
ColumnParallelLinearWithLoRA)
assert isinstance(model.get_submodule("dense2"), RowParallelLinear)
assert isinstance(model.get_submodule("dense2"), RowParallelLinearWithLoRA)
assert isinstance(model.get_submodule("layer1.dense2"),
RowParallelLinearWithLoRA)
@@ -134,8 +130,6 @@ def test_replace_submodules(dist_init, dummy_model):
@pytest.mark.parametrize("device", DEVICES)
def test_lora_model_manager(dist_init, dummy_model, device):
model = dummy_model
model.supported_lora_modules = ["dense1", "dense2", "lm_head"]
model.packed_modules_mapping = {}
model_lora1 = create_lora(1,
model, ["layer1.dense1", "dense2", "lm_head"],
device=device)
@@ -190,13 +184,18 @@ def test_lora_model_manager(dist_init, dummy_model, device):
assert manager.device == device
assert manager.punica_wrapper.device == device
assert hasattr(manager, "supported_lora_modules")
assert sorted(manager.supported_lora_modules) == [
"dense1",
"dense2",
"lm_head",
"output",
]
@pytest.mark.parametrize("device", DEVICES)
def test_lora_lru_cache_model_manager(dist_init, dummy_model, device):
model = dummy_model
model.supported_lora_modules = ["dense1", "dense2", "lm_head"]
model.packed_modules_mapping = {}
model_lora1 = create_lora(1,
model, ["layer1.dense1", "dense2", "lm_head"],
device=device)
@@ -289,8 +288,6 @@ def test_lru_lora_model_manager(dist_init, dummy_model, device):
# This tests just the LRU cache functionality, everything else is
# tested in test_lora_model_manager
model = dummy_model
model.supported_lora_modules = ["dense1", "dense2", "lm_head"]
model.packed_modules_mapping = {}
model_lora1 = create_lora(1,
model, ["layer1.dense1", "dense2", "lm_head"],
device=device)
@@ -572,13 +569,6 @@ def test_worker_adapter_manager(llama_2_7b_model_extra_embeddings,
@pytest.mark.parametrize("device", DEVICES)
def test_packed_loras(dist_init, dummy_model_gate_up, device):
model = dummy_model_gate_up
model.supported_lora_modules = ["gate_up_proj"]
model.packed_modules_mapping = {
"gate_up_proj": [
"gate_proj",
"up_proj",
],
}
model_lora = create_packed_lora(
1,
model,