[Model] Add base class for LoRA-supported models (#5018)

This commit is contained in:
Cyrus Leung
2024-06-27 16:03:04 +08:00
committed by GitHub
parent d12af207d2
commit 96354d6a29
20 changed files with 270 additions and 75 deletions
+10 -2
View File
@@ -26,6 +26,7 @@ from typing import Any, Dict, Iterable, List, Optional, Tuple
import torch
from torch import nn
from transformers import PretrainedConfig
from vllm.attention import Attention, AttentionMetadata
from vllm.config import CacheConfig, LoRAConfig
@@ -51,6 +52,8 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.model_executor.utils import set_weight_attrs
from vllm.sequence import SamplerOutput
from .interfaces import SupportsLoRA
class MiniCPMMoE(nn.Module):
"""A tensor-parallel MoE implementation that shards each expert
@@ -388,7 +391,9 @@ class MiniCPMModel(nn.Module):
return hidden_states
class MiniCPMForCausalLM(nn.Module):
class MiniCPMForCausalLM(nn.Module, SupportsLoRA):
supports_lora = True
packed_modules_mapping = {
"qkv_proj": [
"q_proj",
@@ -418,13 +423,16 @@ class MiniCPMForCausalLM(nn.Module):
def __init__(
self,
config,
config: PretrainedConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
lora_config: Optional[LoRAConfig] = None,
) -> None:
super().__init__()
self.config = config
self.lora_config = lora_config
self.num_experts = getattr(self.config, "num_experts", 0)
self.quant_config = quant_config
self.model = MiniCPMModel(config,