[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
+13 -21
View File
@@ -103,7 +103,6 @@ class LoRAModel:
def check_lora_name(self, lora_name: str) -> bool:
return lora_name in self.loras
# (yard1): TODO see if we can derive target_embedding_padding automatically
@classmethod
def from_lora_tensors(
cls,
@@ -112,9 +111,7 @@ class LoRAModel:
peft_helper: PEFTHelper,
device: str = "cuda",
dtype: torch.dtype | None = None,
target_embedding_padding: int | None = None,
embedding_modules: dict[str, str] | None = None,
embedding_padding_modules: list[str] | None = None,
model_vocab_size: int | None = None,
weights_mapper: WeightsMapper | None = None,
) -> "LoRAModel":
"""Create a LoRAModel from a dictionary of tensors."""
@@ -132,22 +129,21 @@ class LoRAModel:
)
if is_lora_a:
if (
"lora_embedding_A" in tensor_name
and model_vocab_size is not None
and model_vocab_size != tensor.shape[1]
):
raise RuntimeError(
f"The embedding LoRA size({tensor.shape[1]}) must be consistent"
f" with the base model's vocabulary size({model_vocab_size})."
)
loras[module_name].lora_a = tensor.to(device=device, dtype=dtype)
if pin_memory:
loras[module_name].lora_a = loras[module_name].lora_a.pin_memory()
else:
loras[module_name].lora_b = tensor.to(device=device, dtype=dtype)
assert embedding_padding_modules is not None
if (
any(name in module_name for name in embedding_padding_modules)
and target_embedding_padding is not None
):
lora_b = loras[module_name].lora_b
assert target_embedding_padding >= lora_b.shape[0]
addition = target_embedding_padding - lora_b.shape[0]
loras[module_name].lora_b = torch.nn.functional.pad(
lora_b, (0, 0, 0, addition)
)
if pin_memory:
loras[module_name].lora_b = loras[module_name].lora_b.pin_memory()
@@ -163,9 +159,7 @@ class LoRAModel:
lora_model_id: int | None = None,
device: str = "cuda",
dtype: torch.dtype | None = None,
target_embedding_padding: int | None = None,
embedding_modules: dict[str, str] | None = None,
embedding_padding_modules: list[str] | None = None,
model_vocab_size: int | None = None,
weights_mapper: WeightsMapper | None = None,
tensorizer_config_dict: dict | None = None,
) -> "LoRAModel":
@@ -287,9 +281,7 @@ class LoRAModel:
peft_helper=peft_helper,
device=device,
dtype=dtype,
target_embedding_padding=target_embedding_padding,
embedding_modules=embedding_modules,
embedding_padding_modules=embedding_padding_modules,
model_vocab_size=model_vocab_size,
weights_mapper=weights_mapper,
)
+1 -5
View File
@@ -34,12 +34,10 @@ class WorkerLoRAManager:
vllm_config: VllmConfig,
device: torch.device,
embedding_modules: dict[str, str],
embedding_padding_modules: list[str],
lora_model_cls: type[LoRAModel] = LoRAModel,
):
self._lora_model_cls = lora_model_cls
self.embedding_modules = embedding_modules
self.embedding_padding_modules = embedding_padding_modules
self._cached_dummy_lora: None | Literal[False] | LoRAModel = False
self.max_num_seqs = vllm_config.scheduler_config.max_num_seqs
self.max_num_batched_tokens = (
@@ -121,9 +119,7 @@ class WorkerLoRAManager:
lora_model_id=lora_request.lora_int_id,
device="cpu",
dtype=self.lora_config.lora_dtype,
target_embedding_padding=self.vocab_size,
embedding_modules=self.embedding_modules,
embedding_padding_modules=self.embedding_padding_modules,
model_vocab_size=self.vocab_size,
tensorizer_config_dict=lora_request.tensorizer_config_dict,
weights_mapper=hf_to_vllm_mapper,
)