[MoE Refactor] Move the shared/fused expert output sum into MoERunnerBase (#35949)
Signed-off-by: Bill Nell <bnell@redhat.com> Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
@@ -14,7 +14,6 @@ from vllm.config import CacheConfig, ModelConfig, VllmConfig
|
||||
from vllm.distributed import (
|
||||
get_pp_group,
|
||||
get_tensor_model_parallel_world_size,
|
||||
tensor_model_parallel_all_reduce,
|
||||
)
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
@@ -71,7 +70,6 @@ class FusedMoEBlock(nn.Module):
|
||||
top_k=config.moe_top_k,
|
||||
hidden_size=config.hidden_size,
|
||||
intermediate_size=config.moe_intermediate_size,
|
||||
reduce_results=False,
|
||||
renormalize=config.norm_expert_weight,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.experts",
|
||||
@@ -94,8 +92,6 @@ class FusedMoEBlock(nn.Module):
|
||||
final_hidden_states = self.experts(
|
||||
hidden_states=hidden_states, router_logits=router_logits
|
||||
)
|
||||
if self.tp_size > 1:
|
||||
final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
|
||||
|
||||
return final_hidden_states.view(orig_shape)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user