[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
+10 -4
View File
@@ -27,7 +27,7 @@ from torch import nn
from transformers import PretrainedConfig
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_world_size
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.layernorm import RMSNorm
@@ -89,6 +89,7 @@ class XverseAttention(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
@@ -133,7 +134,8 @@ class XverseAttention(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,
@@ -155,6 +157,7 @@ class XverseDecoderLayer(nn.Module):
def __init__(
self,
config: PretrainedConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
@@ -175,6 +178,7 @@ class XverseDecoderLayer(nn.Module):
quant_config=quant_config,
bias=getattr(config, "bias", False),
sliding_window=sliding_window,
cache_config=cache_config,
)
self.mlp = XverseMLP(
hidden_size=self.hidden_size,
@@ -221,6 +225,7 @@ class XverseModel(nn.Module):
def __init__(
self,
config: PretrainedConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
lora_config: Optional[LoRAConfig] = None,
) -> None:
@@ -237,7 +242,7 @@ class XverseModel(nn.Module):
org_num_embeddings=config.vocab_size,
)
self.layers = nn.ModuleList([
XverseDecoderLayer(config, quant_config)
XverseDecoderLayer(config, cache_config, quant_config)
for _ in range(config.num_hidden_layers)
])
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
@@ -295,13 +300,14 @@ class XverseForCausalLM(nn.Module):
def __init__(
self,
config: PretrainedConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
lora_config=None,
) -> None:
super().__init__()
self.config = config
self.quant_config = quant_config
self.model = XverseModel(config, quant_config)
self.model = XverseModel(config, cache_config, quant_config)
self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
self.logits_processor = LogitsProcessor(config.vocab_size)
self.sampler = Sampler()