[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:
@@ -135,7 +135,6 @@ class Llama4MoE(nn.Module):
|
||||
custom_routing_function=Llama4MoE.custom_routing_function,
|
||||
intermediate_size=intermediate_size_moe,
|
||||
apply_router_weight_on_input=True,
|
||||
reduce_results=False,
|
||||
renormalize=False,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.experts",
|
||||
@@ -151,19 +150,14 @@ class Llama4MoE(nn.Module):
|
||||
|
||||
router_logits, _ = self.router(hidden_states)
|
||||
|
||||
shared_out, routed_out = self.experts(
|
||||
experts_out = self.experts(
|
||||
hidden_states=hidden_states,
|
||||
router_logits=router_logits,
|
||||
)
|
||||
experts_out = routed_out + shared_out
|
||||
|
||||
if self.is_sequence_parallel:
|
||||
experts_out = tensor_model_parallel_all_gather(experts_out, 0)
|
||||
experts_out = experts_out[:num_tokens]
|
||||
elif self.tp_size > 1:
|
||||
experts_out = self.experts.maybe_all_reduce_tensor_model_parallel(
|
||||
experts_out
|
||||
)
|
||||
|
||||
return experts_out
|
||||
|
||||
|
||||
Reference in New Issue
Block a user