[MoE] Use GateLinear for all MoE models (#58234)
Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
This commit is contained in:
@@ -32,11 +32,13 @@ from vllm.distributed import (
|
||||
from vllm.distributed.utils import split_tensor_along_last_dim
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoEFactory
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
FusedMoEFactory,
|
||||
GateLinear,
|
||||
)
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
QKVParallelLinear,
|
||||
ReplicatedLinear,
|
||||
RowParallelLinear,
|
||||
)
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
@@ -84,11 +86,9 @@ class OlmoeMoE(nn.Module):
|
||||
self.hidden_size = hidden_size
|
||||
|
||||
# Gate always runs at half / full precision for now.
|
||||
self.gate = ReplicatedLinear(
|
||||
self.gate = GateLinear(
|
||||
hidden_size,
|
||||
num_experts,
|
||||
bias=False,
|
||||
quant_config=None,
|
||||
prefix=f"{prefix}.gate",
|
||||
)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user