Release: 0.1.0

README.md

@@ -90,7 +90,17 @@ from xllm.datasets import GeneralDataset
 from xllm.experiments import Experiment
 
 # Init Config which controls the internal logic of xllm
-config = Config(model_name_or_path="HuggingFaceH4/zephyr-7b-beta")
+# QLoRA example
+config = Config(
+  model_name_or_path="HuggingFaceH4/zephyr-7b-beta",
+  stabilize=True,
+  apply_lora=True,
+  load_in_4bit=True,
+  push_to_hub=True,
+  hub_private_repo=True,
+  hub_model_id="BobaZooba/SupaDupaZephyr-7B-LoRA",
+  save_steps=1_000,
+)
 
 # Prepare the data
 train_data = ["Hello!"] * 100

setup.py

@@ -62,7 +62,7 @@
 # Setup
 setup(
     name="xllm",
-    version="0.0.10",
+    version="0.1.0",
     description="Simple & Cutting Edge LLM Finetuning",
     license_files=["LICENSE"],
     long_description=open("README.md", "r", encoding="utf-8").read(),

src/xllm/__init__.py

@@ -14,7 +14,7 @@
 
 # ruff: noqa: F401
 
-__version__ = "0.0.10"
+__version__ = "0.1.0"
 
 from . import enums, types
 from .cli.fuse import cli_run_fuse

src/xllm/core/config.py

@@ -344,8 +344,8 @@ class Config:
             "help": "Device map for loading the model",
         },
     )
-    prepare_model_for_kbit_training: bool = field(
-        default=True,
+    prepare_model_for_kbit_training: Optional[bool] = field(
+        default=None,
         metadata={
             "help": "Prepare or not for kbit training",
         },
@@ -1069,3 +1069,10 @@ def lora_model_name_or_path_for_fusing(self) -> str:
             return self.lora_model_local_path
         else:
             raise ValueError("Please set lora_hub_model_id or lora_model_local_path for fusing")
+
+    @property
+    def need_to_prepare_model_for_kbit_training(self) -> bool:
+        if self.prepare_model_for_kbit_training is not None:
+            return self.prepare_model_for_kbit_training
+        else:
+            return self.from_gptq or self.load_in_4bit or self.load_in_8bit

src/xllm/core/dependencies.py

@@ -453,7 +453,7 @@ def build_model(
     )
     model.config.pretraining_tp = 1
 
-    if quantization_config is not None and config.prepare_model_for_kbit_training:
+    if quantization_config is not None and config.need_to_prepare_model_for_kbit_training:
         model = prepare_model_for_kbit_training(
             model=model, use_gradient_checkpointing=config.use_gradient_checkpointing
         )

src/xllm/experiments/base.py

@@ -456,7 +456,7 @@ def bnb_quantization(self) -> None:
         )
         self.model.is_loaded_in_4bit = self.config.load_in_4bit
         self.model.is_loaded_in_8bit = self.config.load_in_8bit
-        if self.config.prepare_model_for_kbit_training:
+        if self.config.need_to_prepare_model_for_kbit_training:
             self.model = prepare_model_for_kbit_training(
                 model=self.model, use_gradient_checkpointing=self.config.use_gradient_checkpointing
             )