[Experimental] Add multi-LoRA support (#1804)

Co-authored-by: Chen Shen <scv119@gmail.com>
Co-authored-by: Shreyas Krishnaswamy <shrekris@anyscale.com>
Co-authored-by: Avnish Narayan <avnish@anyscale.com>
This commit is contained in:
Antoni Baum
2024-01-23 15:26:37 -08:00
committed by GitHub
co-authored by Chen Shen Shreyas Krishnaswamy Avnish Narayan
parent 9c1352eb57
commit 9b945daaf1
52 changed files with 8035 additions and 126 deletions
+25 -6
View File
@@ -38,13 +38,14 @@ from vllm.model_executor.layers.linear import (LinearMethodBase,
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.sampler import Sampler
from vllm.model_executor.layers.vocab_parallel_embedding import (
VocabParallelEmbedding, ParallelLMHead)
VocabParallelEmbedding, ParallelLMHead, DEFAULT_VOCAB_PADDING_SIZE)
from vllm.model_executor.parallel_utils.parallel_state import (
get_tensor_model_parallel_world_size)
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.model_executor.weight_utils import (default_weight_loader,
hf_model_weights_iterator)
from vllm.sequence import SamplerOutput
from vllm.config import LoRAConfig
KVCache = Tuple[torch.Tensor, torch.Tensor]
@@ -225,14 +226,19 @@ class LlamaModel(nn.Module):
self,
config: LlamaConfig,
linear_method: Optional[LinearMethodBase] = None,
lora_config: Optional[LoRAConfig] = None,
) -> None:
super().__init__()
self.config = config
self.padding_idx = config.pad_token_id
self.vocab_size = config.vocab_size
lora_vocab = (lora_config.lora_extra_vocab_size *
(lora_config.max_loras or 1)) if lora_config else 0
self.vocab_size = config.vocab_size + lora_vocab
self.org_vocab_size = config.vocab_size
self.embed_tokens = VocabParallelEmbedding(
config.vocab_size,
self.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
)
self.layers = nn.ModuleList([
LlamaDecoderLayer(config, linear_method)
@@ -263,18 +269,31 @@ class LlamaModel(nn.Module):
class LlamaForCausalLM(nn.Module):
supports_lora = True
def __init__(
self,
config: LlamaConfig,
linear_method: Optional[LinearMethodBase] = None,
lora_config: Optional[LoRAConfig] = None,
) -> None:
super().__init__()
self.config = config
self.linear_method = linear_method
self.model = LlamaModel(config, linear_method)
self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
self.sampler = Sampler(config.vocab_size)
self.model = LlamaModel(config, linear_method, lora_config=lora_config)
unpadded_vocab_size = config.vocab_size
if lora_config:
unpadded_vocab_size += lora_config.lora_extra_vocab_size
self.lm_head = ParallelLMHead(
unpadded_vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
padding_size=DEFAULT_VOCAB_PADDING_SIZE
# We need bigger padding if using lora for kernel
# compatibility
if not lora_config else lora_config.lora_vocab_padding_size,
)
self.sampler = Sampler(unpadded_vocab_size, config.vocab_size)
def forward(
self,
+27 -6
View File
@@ -38,13 +38,14 @@ from vllm.model_executor.layers.linear import (LinearMethodBase,
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.sampler import Sampler
from vllm.model_executor.layers.vocab_parallel_embedding import (
VocabParallelEmbedding, ParallelLMHead)
VocabParallelEmbedding, ParallelLMHead, DEFAULT_VOCAB_PADDING_SIZE)
from vllm.model_executor.parallel_utils.parallel_state import (
get_tensor_model_parallel_world_size)
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.model_executor.weight_utils import (default_weight_loader,
hf_model_weights_iterator)
from vllm.sequence import SamplerOutput
from vllm.config import LoRAConfig
KVCache = Tuple[torch.Tensor, torch.Tensor]
@@ -220,15 +221,20 @@ class MistralModel(nn.Module):
self,
config: MistralConfig,
linear_method: Optional[LinearMethodBase] = None,
lora_config: Optional[LoRAConfig] = None,
) -> None:
super().__init__()
self.config = config
self.padding_idx = config.pad_token_id
self.vocab_size = config.vocab_size
lora_vocab = (lora_config.lora_extra_vocab_size *
(lora_config.max_loras or 1)) if lora_config else 0
self.vocab_size = config.vocab_size + lora_vocab
self.org_vocab_size = config.vocab_size
self.embed_tokens = VocabParallelEmbedding(
config.vocab_size,
self.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
)
self.layers = nn.ModuleList([
MistralDecoderLayer(config, linear_method)
@@ -259,18 +265,33 @@ class MistralModel(nn.Module):
class MistralForCausalLM(nn.Module):
supports_lora = True
def __init__(
self,
config: MistralConfig,
linear_method: Optional[LinearMethodBase] = None,
lora_config: Optional[LoRAConfig] = None,
) -> None:
super().__init__()
self.config = config
self.linear_method = linear_method
self.model = MistralModel(config, linear_method)
self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
self.sampler = Sampler(config.vocab_size)
self.model = MistralModel(config,
linear_method,
lora_config=lora_config)
unpadded_vocab_size = config.vocab_size
if lora_config:
unpadded_vocab_size += lora_config.lora_extra_vocab_size
self.lm_head = ParallelLMHead(
unpadded_vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
padding_size=DEFAULT_VOCAB_PADDING_SIZE
# We need bigger padding if using lora for kernel
# compatibility
if not lora_config else lora_config.lora_vocab_padding_size,
)
self.sampler = Sampler(unpadded_vocab_size, config.vocab_size)
def forward(
self,