support for DoRA w/ PEFT (#1363)

requirements.txt

@@ -1,7 +1,7 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 packaging==23.2
-peft @ git+https://github.com/huggingface/peft.git
-transformers @ git+https://github.com/huggingface/transformers.git@ae49b218c3d718df90d8e4a109016450fb8f0632
+peft==0.9.0
+transformers==4.38.2
 tokenizers==0.15.0
 bitsandbytes>=0.41.1
 accelerate==0.26.1

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -178,6 +178,7 @@ class LoraConfig(BaseModel):
     lora_dropout: Optional[float] = None
     peft_layers_to_transform: Optional[List[int]] = None
     peft: Optional[PeftConfig] = None
+    peft_use_dora: Optional[bool] = None
 
     lora_on_cpu: Optional[bool] = None
     gptq: Optional[bool] = None
@@ -233,6 +234,17 @@ def validate_qlora(self):
                     raise ValueError("Require cfg.load_in_4bit to be True for qlora")
         return self
 
+    @model_validator(mode="before")
+    @classmethod
+    def validate_quantized_dora(cls, data):
+        if data.get("peft_use_dora") and (
+            data.get("load_in_8bit") or data.get("load_in_4bit")
+        ):
+            raise ValueError(
+                "`peft_use_dora` is not currently compatible with quantized weights."
+            )
+        return data
+
 
 class ReLoRAConfig(BaseModel):
     """ReLoRA configuration subset"""

src/axolotl/utils/models.py

@@ -830,6 +830,8 @@ def load_lora(model, cfg, inference=False, config_only=False):
     if loftq_bits:
         lora_config_kwargs["loftq_config"] = LoftQConfig(loftq_bits=loftq_bits)
         lora_config_kwargs["init_lora_weights"] = "loftq"
+    if cfg.peft_use_dora:
+        lora_config_kwargs["use_dora"] = cfg.peft_use_dora
 
     lora_config = LoraConfig(
         r=cfg.lora_r,