import os import pathlib import random import string import tempfile import time from concurrent.futures import ThreadPoolExecutor from typing import Iterable, List import gradio as gr import huggingface_hub import torch import yaml from gradio_logsview.logsview import Log, LogsView, LogsViewRunner from mergekit.config import MergeConfiguration from clean_community_org import garbage_collect_empty_models has_gpu = torch.cuda.is_available() # Running directly from Python doesn't work well with Gradio+run_process because of: # Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method # Let's use the CLI instead. # # import mergekit.merge # from mergekit.common import parse_kmb # from mergekit.options import MergeOptions # # merge_options = ( # MergeOptions( # copy_tokenizer=True, # cuda=True, # low_cpu_memory=True, # write_model_card=True, # ) # if has_gpu # else MergeOptions( # allow_crimes=True, # out_shard_size=parse_kmb("1B"), # lazy_unpickle=True, # write_model_card=True, # ) # ) cli = "mergekit-yaml config.yaml merge --copy-tokenizer" + ( " --cuda --low-cpu-memory" if has_gpu else " --allow-crimes --out-shard-size 1B --lazy-unpickle" ) MARKDOWN_DESCRIPTION = """ # mergekit-gui The fastest way to perform a model merge 🔥 Specify a YAML configuration file (see examples below) and a HF token and this app will perform the merge and upload the merged model to your user profile. """ MARKDOWN_ARTICLE = """ ___ ## Merge Configuration [Mergekit](https://github.com/arcee-ai/mergekit) configurations are YAML documents specifying the operations to perform in order to produce your merged model. Below are the primary elements of a configuration file: - `merge_method`: Specifies the method to use for merging models. See [Merge Methods](https://github.com/arcee-ai/mergekit#merge-methods) for a list. - `slices`: Defines slices of layers from different models to be used. This field is mutually exclusive with `models`. - `models`: Defines entire models to be used for merging. This field is mutually exclusive with `slices`. - `base_model`: Specifies the base model used in some merging methods. - `parameters`: Holds various parameters such as weights and densities, which can also be specified at different levels of the configuration. - `dtype`: Specifies the data type used for the merging operation. - `tokenizer_source`: Determines how to construct a tokenizer for the merged model. ## Merge Methods A quick overview of the currently supported merge methods: | Method | `merge_method` value | Multi-Model | Uses base model | | -------------------------------------------------------------------------------------------- | -------------------- | ----------- | --------------- | | Linear ([Model Soups](https://arxiv.org/abs/2203.05482)) | `linear` | ✅ | ❌ | | SLERP | `slerp` | ❌ | ✅ | | [Task Arithmetic](https://arxiv.org/abs/2212.04089) | `task_arithmetic` | ✅ | ✅ | | [TIES](https://arxiv.org/abs/2306.01708) | `ties` | ✅ | ✅ | | [DARE](https://arxiv.org/abs/2311.03099) [TIES](https://arxiv.org/abs/2306.01708) | `dare_ties` | ✅ | ✅ | | [DARE](https://arxiv.org/abs/2311.03099) [Task Arithmetic](https://arxiv.org/abs/2212.04089) | `dare_linear` | ✅ | ✅ | | Passthrough | `passthrough` | ❌ | ❌ | | [Model Stock](https://arxiv.org/abs/2403.19522) | `model_stock` | ✅ | ✅ | ## Citation This GUI is powered by [Arcee's MergeKit](https://arxiv.org/abs/2403.13257). If you use it in your research, please cite the following paper: ``` @article{goddard2024arcee, title={Arcee's MergeKit: A Toolkit for Merging Large Language Models}, author={Goddard, Charles and Siriwardhana, Shamane and Ehghaghi, Malikeh and Meyers, Luke and Karpukhin, Vlad and Benedict, Brian and McQuade, Mark and Solawetz, Jacob}, journal={arXiv preprint arXiv:2403.13257}, year={2024} } ``` This Space is heavily inspired by LazyMergeKit by Maxime Labonne (see [Colab](https://colab.research.google.com/drive/1obulZ1ROXHjYLn6PPZJwRR6GzgQogxxb)). """ examples = [[str(f)] for f in pathlib.Path("examples").glob("*.yml")] # Do not set community token as `HF_TOKEN` to avoid accidentally using it in merge scripts. # `COMMUNITY_HF_TOKEN` is used to upload models to the community organization (https://huggingface.co/mergekit-community) # when user do not provide a token. COMMUNITY_HF_TOKEN = os.getenv("COMMUNITY_HF_TOKEN") def merge(yaml_config: str, hf_token: str, repo_name: str) -> Iterable[List[Log]]: runner = LogsViewRunner() if not yaml_config: yield runner.log("Empty yaml, pick an example below", level="ERROR") return try: merge_config = MergeConfiguration.model_validate(yaml.safe_load(yaml_config)) except Exception as e: yield runner.log(f"Invalid yaml {e}", level="ERROR") return is_community_model = False if not hf_token: if "/" in repo_name and not repo_name.startswith("mergekit-community/"): yield runner.log( f"Cannot upload merge model to namespace {repo_name.split('/')[0]}: you must provide a valid token.", level="ERROR", ) return yield runner.log( "No HF token provided. Your merged model will be uploaded to the https://huggingface.co/mergekit-community organization." ) is_community_model = True if not COMMUNITY_HF_TOKEN: raise gr.Error("Cannot upload to community org: community token not set by Space owner.") hf_token = COMMUNITY_HF_TOKEN api = huggingface_hub.HfApi(token=hf_token) with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname: tmpdir = pathlib.Path(tmpdirname) merged_path = tmpdir / "merged" merged_path.mkdir(parents=True, exist_ok=True) config_path = merged_path / "config.yaml" config_path.write_text(yaml_config) yield runner.log(f"Merge configuration saved in {config_path}") if not repo_name: yield runner.log("No repo name provided. Generating a random one.") repo_name = f"mergekit-{merge_config.merge_method}" # Make repo_name "unique" (no need to be extra careful on uniqueness) repo_name += "-" + "".join(random.choices(string.ascii_lowercase, k=7)) repo_name = repo_name.replace("/", "-").strip("-") if is_community_model and not repo_name.startswith("mergekit-community/"): repo_name = f"mergekit-community/{repo_name}" try: yield runner.log(f"Creating repo {repo_name}") repo_url = api.create_repo(repo_name, exist_ok=True) yield runner.log(f"Repo created: {repo_url}") except Exception as e: yield runner.log(f"Error creating repo {e}", level="ERROR") return # Set tmp HF_HOME to avoid filling up disk Space tmp_env = os.environ.copy() # taken from https://stackoverflow.com/a/4453495 tmp_env["HF_HOME"] = f"{tmpdirname}/.cache" yield from runner.run_command(cli.split(), cwd=merged_path, env=tmp_env) if runner.exit_code != 0: yield runner.log("Merge failed. Deleting repo as no model is uploaded.", level="ERROR") api.delete_repo(repo_url.repo_id) return yield runner.log("Model merged successfully. Uploading to HF.") yield from runner.run_python( api.upload_folder, repo_id=repo_url.repo_id, folder_path=merged_path / "merge", ) yield runner.log(f"Model successfully uploaded to HF: {repo_url.repo_id}") with gr.Blocks() as demo: gr.Markdown(MARKDOWN_DESCRIPTION) with gr.Row(): filename = gr.Textbox(visible=False, label="filename") config = gr.Code(language="yaml", lines=10, label="config.yaml") with gr.Column(): token = gr.Textbox( lines=1, label="HF Write Token", info="https://hf.co/settings/token", type="password", placeholder="Optional. Will upload merged model to MergeKit Community if empty.", ) repo_name = gr.Textbox( lines=1, label="Repo name", placeholder="Optional. Will create a random name if empty.", ) button = gr.Button("Merge", variant="primary") logs = LogsView(label="Terminal output") gr.Examples( examples, fn=lambda s: (s,), run_on_click=True, label="Examples", inputs=[filename], outputs=[config], ) gr.Markdown(MARKDOWN_ARTICLE) button.click(fn=merge, inputs=[config, token, repo_name], outputs=[logs]) # Run garbage collection every hour to keep the community org clean. # Empty models might exists if the merge fails abruptly (e.g. if user leaves the Space). def _garbage_collect_every_hour(): while True: try: garbage_collect_empty_models(token=COMMUNITY_HF_TOKEN) except Exception as e: print("Error running garbage collection", e) time.sleep(3600) pool = ThreadPoolExecutor() pool.submit(_garbage_collect_every_hour) demo.queue(default_concurrency_limit=1).launch()