[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
@@ -22,7 +22,7 @@ from torch import nn
from transformers import GemmaConfig
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.logger import init_logger
from vllm.model_executor.layers.activation import GeluAndMul
@@ -107,6 +107,7 @@ class GemmaAttention(nn.Module):
head_dim: int,
max_position_embeddings: int = 8192,
rope_theta: float = 10000,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None) -> None:
super().__init__()
self.hidden_size = hidden_size
@@ -155,7 +156,8 @@ class GemmaAttention(nn.Module):
self.attn = Attention(self.num_heads,
self.head_dim,
self.scaling,
num_kv_heads=self.num_kv_heads)
num_kv_heads=self.num_kv_heads,
cache_config=cache_config)
def forward(
self,
@@ -177,6 +179,7 @@ class GemmaDecoderLayer(nn.Module):
def __init__(
self,
config: GemmaConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
@@ -188,6 +191,7 @@ class GemmaDecoderLayer(nn.Module):
head_dim=config.head_dim,
max_position_embeddings=config.max_position_embeddings,
rope_theta=config.rope_theta,
cache_config=cache_config,
quant_config=quant_config,
)
self.mlp = GemmaMLP(
@@ -236,6 +240,7 @@ class GemmaModel(nn.Module):
def __init__(
self,
config: GemmaConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
@@ -246,7 +251,7 @@ class GemmaModel(nn.Module):
config.hidden_size,
)
self.layers = nn.ModuleList([
GemmaDecoderLayer(config, quant_config)
GemmaDecoderLayer(config, cache_config, quant_config)
for _ in range(config.num_hidden_layers)
])
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
@@ -309,6 +314,7 @@ class GemmaForCausalLM(nn.Module):
def __init__(
self,
config: GemmaConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
lora_config: Optional[LoRAConfig] = None,
) -> None:
@@ -316,7 +322,7 @@ class GemmaForCausalLM(nn.Module):
super().__init__()
self.config = config
self.quant_config = quant_config
self.model = GemmaModel(config, quant_config)
self.model = GemmaModel(config, cache_config, quant_config)
self.logits_processor = LogitsProcessor(config.vocab_size)
self.sampler = Sampler()