Fix convert_opt_original_pytorch_checkpoint_to_pytorch.py typo (#22526)

`load_checkpoint()` silently fails because `".qkj_proj." in key` is always `False`, but will eventually cause an error at `model.load_state_dict(state_dict)`.

src/transformers/models/opt/convert_opt_original_pytorch_checkpoint_to_pytorch.py

@@ -55,9 +55,9 @@ def load_checkpoint(checkpoint_path):
 
     keys = list(sd.keys())
     for key in keys:
-        if ".qkj_proj." in key:
+        if ".qkv_proj." in key:
             value = sd[key]
-            # We split QKV in seperate Q,K,V
+            # We split QKV in separate Q,K,V
 
             q_name = key.replace(".qkv_proj.", ".q_proj.")
             k_name = key.replace(".qkv_proj.", ".k_proj.")