[Attention] Mamba attention module refactor - Final part (#44857)
This commit is contained in:
@@ -7,7 +7,7 @@ import torch
|
|||||||
from vllm.config import CacheConfig, ModelConfig, get_current_vllm_config
|
from vllm.config import CacheConfig, ModelConfig, get_current_vllm_config
|
||||||
from vllm.distributed import get_tensor_model_parallel_world_size
|
from vllm.distributed import get_tensor_model_parallel_world_size
|
||||||
from vllm.forward_context import ForwardContext, get_forward_context
|
from vllm.forward_context import ForwardContext, get_forward_context
|
||||||
from vllm.model_executor.custom_op import CustomOp
|
from vllm.model_executor.custom_op import PluggableLayer
|
||||||
from vllm.model_executor.layers.linear import (
|
from vllm.model_executor.layers.linear import (
|
||||||
ColumnParallelLinear,
|
ColumnParallelLinear,
|
||||||
MergedColumnParallelLinear,
|
MergedColumnParallelLinear,
|
||||||
@@ -32,8 +32,8 @@ from vllm.v1.attention.backends.short_conv_attn import ShortConvAttentionMetadat
|
|||||||
|
|
||||||
|
|
||||||
# --8<-- [start:short_conv]
|
# --8<-- [start:short_conv]
|
||||||
@CustomOp.register("short_conv")
|
@PluggableLayer.register("short_conv")
|
||||||
class ShortConv(MambaBase, CustomOp):
|
class ShortConv(MambaBase, PluggableLayer):
|
||||||
# --8<-- [end:short_conv]
|
# --8<-- [end:short_conv]
|
||||||
|
|
||||||
def __init__(
|
def __init__(
|
||||||
|
|||||||
Reference in New Issue
Block a user