Create finetuningresearch python package, allowing it to be pip insta…

…lled

Dockerfile

@@ -16,14 +16,14 @@ RUN apt-get update \
     && rm -rf /var/lib/{apt,dpkg,cache,log}
 
 # Copy the current directory contents into the container at /app
-COPY app/requirements.txt requirements.txt
+COPY src/finetuningresearch/requirements.txt requirements.txt
 
 # Install any needed packages specified in requirements.txt
 RUN pip install --upgrade pip && \
     pip install --no-cache-dir -r requirements.txt && \
     rm requirements.txt
 
-COPY app /app
+COPY src/finetuningresearch /app
 
 # Set the working directory in the container to /app
 WORKDIR /app

README.md

@@ -19,3 +19,21 @@ Use a tag versioning by date / user as needed. For example,
 docker build . -t rparundekar/fine_tune_research:20230110_01
 docker push rparundekar/fine_tune_research:20230110_01
 ```
+
+## Library
+To use this finetuning library as a python package, perform a pip install directly from github. This should install all dependencies as well.
+
+```sh
+pip install -v git+https://github.com/shankarg87/training_research@main
+```
+
+then use it normally in your python code.
+
+```python
+from finetuningresearch import execute
+
+config = """
+<config multiline yaml here>
+"""
+execute(config)
+```

pyproject.toml

@@ -1,19 +1,36 @@
 [project]
-name = "Fine-Tuning Reserch"
+name = "finetuningresearch"
 version = "0.1.0"
 description = "Open source research on fine-tuning LLMs"
-authors = ["Rahul Parundekar <[email protected]>", "Shankar Ganesan <[email protected]>"]
-license = "MIT"
+authors = [
+    {name = "Rahul Parundekar", email= "[email protected]" },
+    {name = "Shankar Ganesan", email = "[email protected]" }
+]
 readme = "README.md"
-homepage = "https://aihero.studio"
-repository = "https://github.com/ai-hero/fine_tune_research"
 classifiers = [
     "Programming Language :: Python",
     "Programming Language :: Python :: 3.9",
     "Programming Language :: Python :: 3.10",
     "Programming Language :: Python :: 3.11",
     "License :: OSI Approved :: MIT License",
 ]
+dependencies = [
+    "transformers>=4.35.0",
+    "peft>=0.5.0",
+    "bitsandbytes>=0.41.1",
+    "accelerate>=0.25.0",
+    "trl>=0.7.2",
+    "pydantic-settings>=2.0.3",
+    "scipy>=1.11.3",
+    "PyYAML>=6.0.1",
+    "datasets>=2.14.6",
+    "einops>=0.7.0",
+    "wandb>=0.15.12",
+    "python-dotenv",
+    "minio>=7.2.0",
+    "fire",
+    "types-PyYAML"
+]
 
 [tool.pytest.ini_options]
 addopts = "-vvv"

src/finetuningresearch/__init__.py

@@ -0,0 +1,4 @@
+"""Helps finetune models."""
+from .sft import execute
+
+__all__ = ["execute"]

app/sft.py → src/finetuningresearch/sft.py

@@ -16,7 +16,7 @@
 from trl import SFTTrainer
 from wandb import Table, finish
 
-from utils import DatasetMover, dump_envs, load_config, peft_module_casting_to_bf16
+from .utils import DatasetMover, dump_envs, load_config, peft_module_casting_to_bf16
 
 CHECKPOINT_DIR = "/mnt/checkpoint"
 DATASET_DIR = "/mnt/dataset"
@@ -647,10 +647,11 @@ def save_model(model: Any, tokenizer: Any, config: dict[str, Any]) -> None:
         raise NotImplementedError("S3 support not implemented yet")
 
 
-def main() -> None:
+def execute(config: dict[str, Any] = {}) -> None:
     """Execute the main training loop."""
     dump_envs()
-    config = load_config()
+    if not config:
+        config = load_config()
 
     # Check if "training" is in config or "batch_inference" is in config, but not both.
     if "training" not in config and "batch_inference" not in config:
@@ -694,4 +695,4 @@ def main() -> None:
 
 
 if __name__ == "__main__":
-    Fire(main)
+    Fire(execute)