convert-llama-hf-to-gguf.py : fix attn_q permute

2023-08-21 04:42:09 +02:00 · 2023-08-21 04:42:09 +02:00 · c818c405e0
parent 58bde5c5c1
commit c818c405e0
1 changed files with 3 additions and 1 deletions
--- 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