Add cfg.lora_target_linear

README.md

@@ -232,6 +232,7 @@ lora_target_modules:
 #  - gate_proj
 #  - down_proj
 #  - up_proj
+lora_target_linear: # if true, will target all linear layers
 lora_modules_to_save:
 #  - embed_tokens
 #  - lm_head

src/axolotl/utils/models.py

@@ -364,14 +364,18 @@ def load_lora(model, cfg):
         PeftModel,
     )
 
-    bits = None
-    if cfg.load_in_4bit:
-        bits = 4
-    elif cfg.load_in_8bit:
-        bits = 8
-    linear_names = find_all_linear_names(bits, model)
-    logging.info(f"found linear modules: {repr(linear_names)}")
-    lora_target_modules = list(set(list(cfg.lora_target_modules) + linear_names))
+    lora_target_modules = list(cfg.lora_target_modules)
+
+    if cfg.lora_target_linear:
+        bits = None
+        if cfg.load_in_4bit:
+            bits = 4
+        elif cfg.load_in_8bit:
+            bits = 8
+
+        linear_names = find_all_linear_names(bits, model)
+        logging.info(f"found linear modules: {repr(linear_names)}")
+        lora_target_modules = list(set(lora_target_modules + linear_names))
 
     lora_config = LoraConfig(
         r=cfg.lora_r,