[hotfix] missing w13_weight_fp8 and w2_weight_fp8 in UE8M0 requantization (#12259)

2025-10-28 19:10:38 -07:00
parent 334543ff3b
commit 83087247d1
4 changed files with 30 additions and 30 deletions
@@ -821,8 +821,8 @@ class LongcatFlashForCausalLM(nn.Module):
            experts = layer.mlp.experts
            if isinstance(experts, DeepEPMoE):
                for w in [
-                    experts.w13_weight_fp8,
-                    experts.w2_weight_fp8,
+                    (experts.w13_weight, experts.w13_weight_scale_inv),
+                    (experts.w2_weight, experts.w2_weight_scale_inv),
                ]:
                    requant_weight_ue8m0_inplace(w[0], w[1], weight_block_size)