[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
+12 -4
View File
@@ -23,6 +23,7 @@ from torch import nn
from transformers import GPTNeoXConfig
from vllm.attention import Attention, AttentionMetadata
from vllm.config import CacheConfig
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
@@ -45,6 +46,7 @@ class GPTNeoXAttention(nn.Module):
def __init__(
self,
config: GPTNeoXConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
):
super().__init__()
@@ -84,7 +86,10 @@ class GPTNeoXAttention(nn.Module):
max_position=max_position_embeddings,
base=rope_theta,
)
self.attn = Attention(self.num_heads, self.head_size, scaling)
self.attn = Attention(self.num_heads,
self.head_size,
scaling,
cache_config=cache_config)
def forward(
self,
@@ -134,6 +139,7 @@ class GPTNeoXLayer(nn.Module):
def __init__(
self,
config: GPTNeoXConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
):
super().__init__()
@@ -142,7 +148,7 @@ class GPTNeoXLayer(nn.Module):
eps=config.layer_norm_eps)
self.post_attention_layernorm = nn.LayerNorm(config.hidden_size,
eps=config.layer_norm_eps)
self.attention = GPTNeoXAttention(config, quant_config)
self.attention = GPTNeoXAttention(config, cache_config, quant_config)
self.mlp = GPTNeoXMLP(config, quant_config)
def forward(
@@ -182,6 +188,7 @@ class GPTNeoXModel(nn.Module):
def __init__(
self,
config: GPTNeoXConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
):
super().__init__()
@@ -192,7 +199,7 @@ class GPTNeoXModel(nn.Module):
config.hidden_size,
)
self.layers = nn.ModuleList([
GPTNeoXLayer(config, quant_config)
GPTNeoXLayer(config, cache_config, quant_config)
for _ in range(config.num_hidden_layers)
])
self.final_layer_norm = nn.LayerNorm(config.hidden_size,
@@ -223,12 +230,13 @@ class GPTNeoXForCausalLM(nn.Module):
def __init__(
self,
config,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
):
super().__init__()
self.config = config
self.quant_config = quant_config
self.gpt_neox = GPTNeoXModel(config, quant_config)
self.gpt_neox = GPTNeoXModel(config, cache_config, quant_config)
self.embed_out = ParallelLMHead(
config.vocab_size,
config.hidden_size,