biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Misc] Enhance attention selector (#4751)

Browse Source
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
Download Patch File Download Diff File Expand all files Collapse all files

15
vllm/model_executor/models/qwen.py
View File

@@ -11,6 +11,7 @@ from torch import nn
from transformers import PretrainedConfig
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 SiluAndMul
from vllm.model_executor.layers.layernorm import RMSNorm
@@ -68,6 +69,7 @@ class QWenAttention(nn.Module):
max_position_embeddings: int,
rope_theta: float = 10000,
rope_scaling: Optional[Dict[str, Any]] = None,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
):
super().__init__()
@@ -101,7 +103,10 @@ class QWenAttention(nn.Module):
base=rope_theta,
rope_scaling=rope_scaling,
)
self.attn = Attention(self.num_heads, self.head_dim, self.scaling)
self.attn = Attention(self.num_heads,
self.head_dim,
self.scaling,
cache_config=cache_config)
def forward(
self,
@@ -123,6 +128,7 @@ class QWenBlock(nn.Module):
def __init__(
self,
config: PretrainedConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
):
super().__init__()
@@ -135,6 +141,7 @@ class QWenBlock(nn.Module):
config.max_position_embeddings,
rope_theta=rope_theta,
rope_scaling=rope_scaling,
cache_config=cache_config,
quant_config=quant_config)
self.ln_2 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
@@ -175,6 +182,7 @@ class QWenModel(nn.Module):
def __init__(
self,
config: PretrainedConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
):
super().__init__()
@@ -186,7 +194,7 @@ class QWenModel(nn.Module):
config.hidden_size,
)
self.h = nn.ModuleList([
QWenBlock(config, quant_config)
QWenBlock(config, cache_config, quant_config)
for _ in range(config.num_hidden_layers)
])
self.ln_f = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
@@ -218,12 +226,13 @@ class QWenLMHeadModel(nn.Module):
def __init__(
self,
config: PretrainedConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
):
super().__init__()
self.config = config
self.quant_config = quant_config
self.transformer = QWenModel(config, quant_config)
self.transformer = QWenModel(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()