biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[LoRA] Continue optimizing MoE LoRA weight loading (#29322)

Browse Source 
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Jee Jee Li
2025-11-27 21:56:28 +08:00
committed by GitHub
parent cf348c8d27
commit 2f5f9acd55
15 changed files with 228 additions and 157 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
tests/lora/test_lora_checkpoints.py
View File

@@ -28,12 +28,13 @@ def test_load_checkpoints(
packed_modules_mapping = BaiChuanBaseForCausalLM.packed_modules_mapping
embedding_modules = BaiChuanBaseForCausalLM.embedding_modules
embed_padding_modules = BaiChuanBaseForCausalLM.embedding_padding_modules
expected_lora_modules: list[str] = []
expected_lora_lst: list[str] = []
for module in BAICHUAN_LORA_MODULES:
if module in packed_modules_mapping:
expected_lora_modules.extend(packed_modules_mapping[module])
expected_lora_lst.extend(packed_modules_mapping[module])
else:
expected_lora_modules.append(module)
expected_lora_lst.append(module)
expected_lora_modules = set(expected_lora_lst)
if lora_name == "baichuan7B":
peft_helper = PEFTHelper.from_local_dir(
baichuan_lora_files, max_position_embeddings=4096
@@ -103,13 +104,13 @@ 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_modules: list[str] = []
expected_lora_lst: list[str] = []
for module in BAICHUAN_LORA_MODULES:
if module in packed_modules_mapping:
expected_lora_modules.extend(packed_modules_mapping[module])
expected_lora_lst.extend(packed_modules_mapping[module])
else:
expected_lora_modules.append(module)
expected_lora_lst.append(module)
expected_lora_modules = set(expected_lora_lst)
hf_to_vllm_mapper = WeightsMapper(
orig_to_new_prefix={
"model.": "language_model.model.",

8
tests/lora/test_lora_huggingface.py
View File

@@ -26,13 +26,13 @@ def test_load_checkpoints_from_huggingface(lora_fixture_name, request):
packed_modules_mapping = LlamaForCausalLM.packed_modules_mapping
embedding_modules = LlamaForCausalLM.embedding_modules
embed_padding_modules = LlamaForCausalLM.embedding_padding_modules
expected_lora_modules: list[str] = []
expected_lora_lst: list[str] = []
for module in LLAMA_LORA_MODULES:
if module in packed_modules_mapping:
expected_lora_modules.extend(packed_modules_mapping[module])
expected_lora_lst.extend(packed_modules_mapping[module])
else:
expected_lora_modules.append(module)
expected_lora_lst.append(module)
expected_lora_modules = set(expected_lora_lst)
lora_path = get_adapter_absolute_path(lora_name)
# lora loading should work for either absolute path and huggingface id.