[chore] clean-up weight prepack for INT8 MoE (#50116)
Signed-off-by: Fadi Arafeh <fadi.arafeh@arm.com>
This commit is contained in:
+1
@@ -162,6 +162,7 @@ class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
|
||||
routing_tables=layer._expert_routing_tables(),
|
||||
layer=layer,
|
||||
)
|
||||
self.moe_kernel.fused_experts.process_weights_after_loading(layer)
|
||||
|
||||
def maybe_make_prepare_finalize(
|
||||
self,
|
||||
|
||||
@@ -114,6 +114,7 @@ class Int8OnlineMoEMethod(OnlineMoEMethodBase):
|
||||
routing_tables=layer._expert_routing_tables(),
|
||||
layer=layer,
|
||||
)
|
||||
self.moe_kernel.fused_experts.process_weights_after_loading(layer)
|
||||
|
||||
def get_fused_moe_quant_config(
|
||||
self, layer: torch.nn.Module
|
||||
|
||||
Reference in New Issue
Block a user