From 4bdd22ea771120ea64cc1fcc7c0233e5f468a0ad Mon Sep 17 00:00:00 2001
From: turboderp <11859846+turboderp@users.noreply.github.com>
Date: Sun, 1 Mar 2026 03:13:03 +0100
Subject: [PATCH] BlockSparseMLP: Make sure bias is always applied during
 calibration

---
 exllamav3/modules/block_sparse_mlp.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/exllamav3/modules/block_sparse_mlp.py b/exllamav3/modules/block_sparse_mlp.py
index 35fea8b..6da0a59 100644
--- a/exllamav3/modules/block_sparse_mlp.py
+++ b/exllamav3/modules/block_sparse_mlp.py
@@ -113,6 +113,8 @@ def routing_dots(bsz, cfg, y, params):
         activate_all_experts = params.get("activate_all_experts")
         if activate_all_experts:
             routing_weights = router_logits.sigmoid()
+            if cfg.e_score_correction_bias is not None:
+                routing_weights += cfg.e_score_correction_bias.unsqueeze(0)
             factor = cfg.routed_scaling_factor / (routing_weights.sum(dim = -1, keepdim = True) + 1e-20)
             routing_weights *= factor
             selected_experts = (