[Misc] Enhance attention selector (#4751)

This commit is contained in:
Woosuk Kwon
2024-05-13 10:47:25 -07:00
committed by GitHub
parent e7c46b9527
commit 0fca3cdcf2
49 changed files with 573 additions and 220 deletions
+13 -4
View File
@@ -28,7 +28,7 @@ from torch import nn
from transformers import LlamaConfig
from vllm.attention import Attention, AttentionMetadata
from vllm.config import LoRAConfig
from vllm.config import CacheConfig, LoRAConfig
from vllm.distributed import (get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size)
from vllm.model_executor.layers.activation import SiluAndMul
@@ -94,6 +94,7 @@ class LlamaAttention(nn.Module):
quant_config: Optional[QuantizationConfig] = None,
bias: bool = False,
sliding_window: Optional[int] = None,
cache_config: Optional[CacheConfig] = None,
) -> None:
super().__init__()
self.hidden_size = hidden_size
@@ -153,7 +154,8 @@ class LlamaAttention(nn.Module):
self.head_dim,
self.scaling,
num_kv_heads=self.num_kv_heads,
sliding_window=sliding_window)
sliding_window=sliding_window,
cache_config=cache_config)
def forward(
self,
@@ -176,6 +178,7 @@ class LlamaDecoderLayer(nn.Module):
def __init__(
self,
config: LlamaConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
@@ -204,6 +207,7 @@ class LlamaDecoderLayer(nn.Module):
quant_config=quant_config,
bias=attention_bias,
sliding_window=sliding_window,
cache_config=cache_config,
)
self.mlp = LlamaMLP(
hidden_size=self.hidden_size,
@@ -251,6 +255,7 @@ class LlamaModel(nn.Module):
def __init__(
self,
config: LlamaConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
lora_config: Optional[LoRAConfig] = None,
) -> None:
@@ -267,7 +272,7 @@ class LlamaModel(nn.Module):
org_num_embeddings=config.vocab_size,
)
self.layers = nn.ModuleList([
LlamaDecoderLayer(config, quant_config)
LlamaDecoderLayer(config, cache_config, quant_config)
for _ in range(config.num_hidden_layers)
])
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
@@ -332,12 +337,16 @@ class LlamaForCausalLM(nn.Module):
def __init__(
self,
config: LlamaConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
lora_config: Optional[LoRAConfig] = None,
) -> None:
super().__init__()
self.config = config
self.model = LlamaModel(config, quant_config, lora_config=lora_config)
self.model = LlamaModel(config,
cache_config,
quant_config,
lora_config=lora_config)
self.unpadded_vocab_size = config.vocab_size
if lora_config:
self.unpadded_vocab_size += lora_config.lora_extra_vocab_size