[Misc] Enhance attention selector (#4751)
This commit is contained in:
@@ -28,7 +28,7 @@ import torch
|
||||
from torch import nn
|
||||
|
||||
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_rank,
|
||||
get_tensor_model_parallel_world_size,
|
||||
tensor_model_parallel_all_reduce)
|
||||
@@ -181,6 +181,7 @@ class MiniCPMAttention(nn.Module):
|
||||
rope_theta: float = 10000,
|
||||
rope_scaling: Optional[Dict[str, Any]] = None,
|
||||
max_position_embeddings: int = 8192,
|
||||
cache_config: Optional[CacheConfig] = None,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
@@ -234,7 +235,8 @@ class MiniCPMAttention(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,
|
||||
@@ -259,6 +261,7 @@ class MiniCPMDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config,
|
||||
cache_config: Optional[CacheConfig] = None,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
@@ -275,6 +278,7 @@ class MiniCPMDecoderLayer(nn.Module):
|
||||
rope_theta=rope_theta,
|
||||
rope_scaling=rope_scaling,
|
||||
max_position_embeddings=max_position_embeddings,
|
||||
cache_config=cache_config,
|
||||
quant_config=quant_config,
|
||||
)
|
||||
self.num_experts = getattr(self.config, "num_experts", 0)
|
||||
@@ -330,6 +334,7 @@ class MiniCPMModel(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config,
|
||||
cache_config: Optional[CacheConfig] = None,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
lora_config: Optional[LoRAConfig] = None,
|
||||
) -> None:
|
||||
@@ -346,7 +351,7 @@ class MiniCPMModel(nn.Module):
|
||||
org_num_embeddings=config.vocab_size,
|
||||
)
|
||||
self.layers = nn.ModuleList([
|
||||
MiniCPMDecoderLayer(config, quant_config)
|
||||
MiniCPMDecoderLayer(config, cache_config, quant_config)
|
||||
for _ in range(config.num_hidden_layers)
|
||||
])
|
||||
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
@@ -413,6 +418,7 @@ class MiniCPMForCausalLM(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config,
|
||||
cache_config: Optional[CacheConfig] = None,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
lora_config: Optional[LoRAConfig] = None,
|
||||
) -> None:
|
||||
@@ -421,6 +427,7 @@ class MiniCPMForCausalLM(nn.Module):
|
||||
self.num_experts = getattr(self.config, "num_experts", 0)
|
||||
self.quant_config = quant_config
|
||||
self.model = MiniCPMModel(config,
|
||||
cache_config,
|
||||
quant_config,
|
||||
lora_config=lora_config)
|
||||
unpadded_vocab_size = config.vocab_size
|
||||
|
||||
Reference in New Issue
Block a user