support lora when kv_channel != hidden_size / num_heads (#9644)

Co-authored-by: Ao Tang <aot@nvidia.com>
NVIDIA · Jul 8, 2024 · f790741 · f790741
1 parent 66c960e
commit f790741
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/nemo/collections/nlp/parts/peft_config.py b/nemo/collections/nlp/parts/peft_config.py
@@ -170,7 +170,7 @@ def __init__(self, cfg):
 
             elif module == PEFT_MODULE_MAP["dense_module"]:
                 adapter_cfg = self._create_lora_config(
-                    cfg, lora_cfg, cfg.hidden_size, cfg.hidden_size, LoraDenseAttentionAdapterConfig
+                    cfg, lora_cfg, projection_size, cfg.hidden_size, LoraDenseAttentionAdapterConfig
                 )
                 name_key_to_cfg[AdapterName.LORA_DENSE_ATTENTION_ADAPTER] = adapter_cfg
                 name_key_to_mcore_mixins[AdapterName.LORA_DENSE_ATTENTION_ADAPTER] = [