[Model] Add base class for LoRA-supported models (#5018)
This commit is contained in:
@@ -54,6 +54,8 @@ from vllm.model_executor.utils import set_weight_attrs
|
||||
from vllm.sequence import SamplerOutput
|
||||
from vllm.utils import print_warning_once
|
||||
|
||||
from .interfaces import SupportsLoRA
|
||||
|
||||
|
||||
class MixtralMoE(nn.Module):
|
||||
"""A tensor-parallel MoE implementation for Mixtral that shards each expert
|
||||
@@ -472,7 +474,9 @@ class MixtralModel(nn.Module):
|
||||
return hidden_states
|
||||
|
||||
|
||||
class MixtralForCausalLM(nn.Module):
|
||||
class MixtralForCausalLM(nn.Module, SupportsLoRA):
|
||||
supports_lora = True
|
||||
|
||||
fall_back_to_pt_during_load = False
|
||||
|
||||
packed_modules_mapping = {
|
||||
@@ -504,7 +508,10 @@ class MixtralForCausalLM(nn.Module):
|
||||
lora_config: Optional[LoRAConfig] = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
|
||||
self.config = config
|
||||
self.lora_config = lora_config
|
||||
|
||||
self.model = MixtralModel(config,
|
||||
cache_config,
|
||||
quant_config,
|
||||
|
||||
Reference in New Issue
Block a user