Unverified Commit e5c78956 authored by Alexander Matveev's avatar Alexander Matveev Committed by GitHub
Browse files

[Bugfix] Fix incorrect use of hidden_states for shared_experts due to...


[Bugfix] Fix incorrect use of hidden_states for shared_experts due to do_naive_dispatch_combine (#28740)
Signed-off-by: default avatarAlexander Matveev <amatveev@redhat.com>
parent 2e0ad629
...@@ -1749,14 +1749,16 @@ class FusedMoE(CustomOp): ...@@ -1749,14 +1749,16 @@ class FusedMoE(CustomOp):
with sp_ctx: with sp_ctx:
if do_naive_dispatch_combine: if do_naive_dispatch_combine:
hidden_states, router_logits = get_ep_group().dispatch( hidden_states_combined, router_logits = get_ep_group().dispatch(
hidden_states, router_logits, self.is_sequence_parallel hidden_states, router_logits, self.is_sequence_parallel
) )
# Matrix multiply. # Matrix multiply.
final_hidden_states = self.quant_method.apply( final_hidden_states = self.quant_method.apply(
layer=self, layer=self,
x=hidden_states, x=hidden_states_combined
if do_naive_dispatch_combine
else hidden_states,
router_logits=router_logits, router_logits=router_logits,
top_k=self.top_k, top_k=self.top_k,
renormalize=self.renormalize, renormalize=self.renormalize,
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment