[fix](gpt-oss): fix quark quantized model in moe bias (#787)

Co-authored-by: perzhang <perzhang@amd.com>

Author: 2 people

atom/model_ops/moe.py

@@ -756,7 +756,7 @@ def create_weights(
 
         if layer.has_bias:
             w13_bias = atom_parameter(
-                torch.empty(
+                torch.zeros(
                     num_experts,
                     2 * intermediate_size_per_partition_after_pad,
                     dtype=torch.bfloat16,
@@ -793,7 +793,7 @@ def create_weights(
 
         if layer.has_bias:
             w2_bias = atom_parameter(
-                torch.empty(
+                torch.zeros(
                     num_experts,
                     hidden_size,
                     dtype=torch.bfloat16,