[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
+21 -10
View File
@@ -30,6 +30,7 @@ from torch import nn
from transformers import MixtralConfig
from vllm.attention import Attention, AttentionMetadata
from vllm.config import CacheConfig
from vllm.distributed import (get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_reduce)
@@ -157,14 +158,17 @@ class MixtralMoE(nn.Module):
class MixtralAttention(nn.Module):
def __init__(self,
hidden_size: int,
num_heads: int,
num_kv_heads: int,
max_position: int = 4096 * 32,
rope_theta: float = 10000,
quant_config: Optional[QuantizationConfig] = None,
sliding_window: Optional[int] = None) -> None:
def __init__(
self,
hidden_size: int,
num_heads: int,
num_kv_heads: int,
max_position: int = 4096 * 32,
rope_theta: float = 10000,
quant_config: Optional[QuantizationConfig] = None,
sliding_window: Optional[int] = None,
cache_config: Optional[CacheConfig] = None,
) -> None:
super().__init__()
self.hidden_size = hidden_size
tp_size = get_tensor_model_parallel_world_size()
@@ -215,6 +219,7 @@ class MixtralAttention(nn.Module):
self.scaling,
num_kv_heads=self.num_kv_heads,
sliding_window=self.sliding_window,
cache_config=cache_config,
)
def forward(
@@ -237,6 +242,7 @@ class MixtralDecoderLayer(nn.Module):
def __init__(
self,
config: MixtralConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
@@ -250,6 +256,7 @@ class MixtralDecoderLayer(nn.Module):
num_kv_heads=config.num_key_value_heads,
rope_theta=rope_theta,
sliding_window=config.sliding_window,
cache_config=cache_config,
quant_config=quant_config)
self.block_sparse_moe = MixtralMoE(config=config,
quant_config=quant_config)
@@ -292,6 +299,7 @@ class MixtralModel(nn.Module):
def __init__(
self,
config: MixtralConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
@@ -303,7 +311,9 @@ class MixtralModel(nn.Module):
config.hidden_size,
)
self.layers = nn.ModuleList([
MixtralDecoderLayer(config, quant_config=quant_config)
MixtralDecoderLayer(config,
cache_config,
quant_config=quant_config)
for _ in range(config.num_hidden_layers)
])
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
@@ -332,12 +342,13 @@ class MixtralForCausalLM(nn.Module):
def __init__(
self,
config: MixtralConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
self.config = config
self.quant_config = quant_config
self.model = MixtralModel(config, quant_config)
self.model = MixtralModel(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()