[LoRA] Cleanup LoRA unused code (#29611)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>

vllm/lora/worker_manager.py

@@ -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,
             )

vllm/model_executor/models/apertus.py

@@ -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,

vllm/model_executor/models/bamba.py

@@ -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(

vllm/model_executor/models/exaone.py

@@ -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__()

vllm/model_executor/models/exaone4.py

@@ -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__()

vllm/model_executor/models/falcon_h1.py

@@ -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(

vllm/model_executor/models/granite.py

@@ -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__()

vllm/model_executor/models/granitemoe.py

@@ -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__()

vllm/model_executor/models/granitemoehybrid.py

@@ -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(

vllm/model_executor/models/granitemoeshared.py

@@ -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__()

vllm/model_executor/models/interfaces.py

@@ -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))
 

vllm/model_executor/models/jamba.py

@@ -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

vllm/model_executor/models/lfm2.py

@@ -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(

vllm/model_executor/models/lfm2_moe.py

@@ -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(

vllm/model_executor/models/llama.py

@@ -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

vllm/model_executor/models/minicpm.py

@@ -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__()

vllm/model_executor/models/minicpm_eagle.py

@@ -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__()

vllm/model_executor/models/minicpmv.py

@@ -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)

vllm/model_executor/models/mixtral.py

@@ -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__()

vllm/model_executor/models/nemotron.py

@@ -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__()