[Bugfix] Disable shared expert overlap if Marlin MoE is used (#28410)
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
@@ -424,6 +424,7 @@ class AWQMoEMethod(FusedMoEMethodBase):
|
||||
if self.quant_config.weight_bits != 4:
|
||||
raise ValueError("AWQMoEMethod only supports 4bit now.")
|
||||
self.quant_type = scalar_types.uint4
|
||||
self.use_marlin = True
|
||||
|
||||
def create_weights(
|
||||
self,
|
||||
|
||||
Reference in New Issue
Block a user