Fixing the other pathways.

huggingface · Aug 14, 2024 · 5c598cc · 5c598cc
1 parent c904766
commit 5c598cc
Showing 1 changed file with 13 additions and 11 deletions.
diff --git a/server/text_generation_server/layers/gptq/exllamav2.py b/server/text_generation_server/layers/gptq/exllamav2.py
@@ -115,29 +115,31 @@ def ext_make_q_matrix(
                 w.qweight,
                 extra.q_perm,
                 extra.q_invperm,
-                none_tensor,
-                none_tensor,
-                none_tensor,
-                none_tensor,
+                none_tensor,  # q_scale
+                none_tensor,  # q_scale_max
+                none_tensor,  # q_groups
+                none_tensor,  # q_group_map
                 w.qzeros,
                 w.scales,
                 w.g_idx.cpu(),
+                none_tensor,  # bias
                 temp_dq,
                 max_dq_rows,
             )
         # GPTQ without g_idx
         else:
             return make_q_matrix(
                 w.qweight,
-                none_tensor,
-                none_tensor,
-                none_tensor,
-                none_tensor,
-                none_tensor,
-                none_tensor,
+                none_tensor,  # q_perm
+                none_tensor,  # q_invperm
+                none_tensor,  # q_scale
+                none_tensor,  # q_scale_max
+                none_tensor,  # q_groups
+                none_tensor,  # q_group_map
                 w.qzeros,
                 w.scales,
-                none_tensor,
+                none_tensor,  # g_idx
+                none_tensor,  # bias
                 temp_dq,
                 max_dq_rows,
             )