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

[LoRA] Cleanup LoRA unused code (#29611)

Browse Source 
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
2025-11-29 14:52:58 +08:00
committed by GitHub
parent 4a80ad0a25
commit 39e63dec7c
46 changed files with 126 additions and 173 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
vllm/model_executor/models/apertus.py
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
vllm/model_executor/models/bamba.py
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
vllm/model_executor/models/exaone.py
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
vllm/model_executor/models/exaone4.py
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
vllm/model_executor/models/falcon_h1.py
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
vllm/model_executor/models/granite.py
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
vllm/model_executor/models/granitemoe.py
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__()

1
vllm/model_executor/models/granitemoehybrid.py
View File

@@ -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(

1
vllm/model_executor/models/granitemoeshared.py
View File

@@ -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
vllm/model_executor/models/interfaces.py
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
vllm/model_executor/models/jamba.py
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
vllm/model_executor/models/lfm2.py
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
vllm/model_executor/models/lfm2_moe.py
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
vllm/model_executor/models/llama.py
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
vllm/model_executor/models/minicpm.py
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__()

1
vllm/model_executor/models/minicpm_eagle.py
View File

@@ -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
vllm/model_executor/models/minicpmv.py
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
vllm/model_executor/models/mixtral.py
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
vllm/model_executor/models/nemotron.py
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
vllm/model_executor/models/nemotron_h.py
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(

1
vllm/model_executor/models/nemotron_nas.py
View File

@@ -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
vllm/model_executor/models/phimoe.py
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
vllm/model_executor/models/solar.py
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__()

1
vllm/model_executor/models/transformers/base.py
View File

@@ -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={