From c818c405e025a766c0ab12afb18c4baec7810a6d Mon Sep 17 00:00:00 2001
From: klosax <131523366+klosax@users.noreply.github.com>
Date: Mon, 21 Aug 2023 04:42:09 +0200
Subject: [PATCH] convert-llama-hf-to-gguf.py : fix attn_q permute

---
 convert-llama-hf-to-gguf.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/convert-llama-hf-to-gguf.py b/convert-llama-hf-to-gguf.py
index 9289efd28..6ee5fdf76 100644
--- a/convert-llama-hf-to-gguf.py
+++ b/convert-llama-hf-to-gguf.py
@@ -264,7 +264,9 @@ for part_name in part_names:
         data = data.squeeze().numpy()
 
         # reverse permute these
-        if name.endswith(".q_proj.weight") or name.endswith(".k_proj.weight"):
+        if name.endswith(".q_proj.weight"):
+            data = reverse_hf_permute(data, head_count)
+        if name.endswith(".k_proj.weight"):
             data = reverse_hf_permute(data, head_count, head_count_kv)
 
         # map tensor names