[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,
)
-1
View File
@@ -482,7 +482,6 @@ class ApertusForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
def __init__(
self,
-1
View File
@@ -419,7 +419,6 @@ class BambaForCausalLM(
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
@classmethod
def get_mamba_state_dtype_from_config(
-1
View File
@@ -457,7 +457,6 @@ class ExaoneForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
"wte": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
-1
View File
@@ -450,7 +450,6 @@ class Exaone4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
-1
View File
@@ -510,7 +510,6 @@ class FalconH1ForCausalLM(
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
@classmethod
def get_mamba_state_dtype_from_config(
-1
View File
@@ -400,7 +400,6 @@ class GraniteForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
-1
View File
@@ -497,7 +497,6 @@ class GraniteMoeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -601,7 +601,6 @@ class GraniteMoeHybridForCausalLM(
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
@classmethod
def get_mamba_state_dtype_from_config(
@@ -263,7 +263,6 @@ class GraniteMoeSharedForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
-3
View File
@@ -347,7 +347,6 @@ class SupportsLoRA(Protocol):
# The `embedding_module` and `embedding_padding_modules`
# are empty by default.
embedding_modules: ClassVar[dict[str, str]] = {}
embedding_padding_modules: ClassVar[list[str]] = []
packed_modules_mapping: dict[str, list[str]] = {}
@@ -359,7 +358,6 @@ class _SupportsLoRAType(Protocol):
packed_modules_mapping: dict[str, list[str]]
embedding_modules: dict[str, str]
embedding_padding_modules: list[str]
@overload
@@ -379,7 +377,6 @@ def supports_lora(
lora_attrs = (
"packed_modules_mapping",
"embedding_modules",
"embedding_padding_modules",
)
missing_attrs = tuple(attr for attr in lora_attrs if not hasattr(model, attr))
-1
View File
@@ -480,7 +480,6 @@ class JambaForCausalLM(
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
config = vllm_config.model_config.hf_config
-1
View File
@@ -422,7 +422,6 @@ class Lfm2ForCausalLM(
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
@classmethod
def get_mamba_state_dtype_from_config(
-1
View File
@@ -602,7 +602,6 @@ class Lfm2MoeForCausalLM(
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
@classmethod
def get_mamba_state_dtype_from_config(
-1
View File
@@ -528,7 +528,6 @@ class LlamaForCausalLM(
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
# Mistral/Llama models can also be loaded with --load-format mistral
# from consolidated.safetensors checkpoints
-1
View File
@@ -568,7 +568,6 @@ class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -305,7 +305,6 @@ class EagleMiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
-1
View File
@@ -1741,5 +1741,4 @@ class MiniCPMV(MiniCPMVBaseModel, SupportsMultiModal, SupportsLoRA):
# so update values before init is called
cls.packed_modules_mapping.update(instance_cls.packed_modules_mapping)
cls.embedding_modules.update(instance_cls.embedding_modules)
cls.embedding_padding_modules += instance_cls.embedding_padding_modules
return instance_cls(vllm_config=vllm_config, prefix=prefix)
-1
View File
@@ -496,7 +496,6 @@ class MixtralForCausalLM(nn.Module, SupportsLoRA, SupportsPP, MixtureOfExperts):
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
-1
View File
@@ -439,7 +439,6 @@ class NemotronForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
-1
View File
@@ -713,7 +713,6 @@ class NemotronHForCausalLM(
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
@classmethod
def get_mamba_state_dtype_from_config(
@@ -387,7 +387,6 @@ class DeciLMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, HasNoOps):
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
# Mistral/Llama models can also be loaded with --load-format mistral
# from consolidated.safetensors checkpoints
-1
View File
@@ -617,7 +617,6 @@ class PhiMoEForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
-1
View File
@@ -426,7 +426,6 @@ class SolarForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
"embed_tokens": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -93,7 +93,6 @@ ALL_ATTENTION_FUNCTIONS["vllm"] = vllm_flash_attention_forward
class Base(nn.Module, VllmModel, SupportsQuant, SupportsLoRA, SupportsPP):
embedding_padding_modules = ["lm_head"]
embedding_modules = ["embed_tokens"] # TODO transformers will have a util to get it
hf_to_vllm_mapper = WeightsMapper(
orig_to_new_prefix={
@@ -43,7 +43,6 @@ class LoRAModelRunnerMixin:
vllm_config,
device,
model.embedding_modules,
model.embedding_padding_modules,
)
return self.lora_manager.create_lora_manager(model)