Saving, Loading, and push_to_hub

Symmetry: save_pretrained / from_pretrained

Every PreTrainedModel, PreTrainedTokenizer, and processor in the library has the same pair: save_pretrained(path) writes config + weights + tokenizer files to a directory, and from_pretrained(path or repo_id) loads them. Local paths and Hub repo ids are interchangeable; the loader sniffs which one you passed.

This symmetry is the reason the entire ecosystem composes. A LoRA trainer outputs a directory; you pass that directory to AutoPeftModelForCausalLM; you pass the merged directory to push_to_hub(); the next person calls from_pretrained on your repo id. End-to-end: same two methods.

push_to_hub: the underrated half

Both model.push_to_hub("yourname/your-repo") and tokenizer.push_to_hub(...) create or update a Hub repo. The first call creates with a default README; subsequent calls update. create_repo() from huggingface_hub gives you the explicit control (private, exist_ok, organization, etc.).

For weight uploads, prefer the safetensors format (default since transformers 4.34). It's safer (no pickle), faster (memory-mapped), and smaller (no Python overhead).

Code

Save locally then push to Hub·python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "Qwen/Qwen2.5-1.5B-Instruct"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo)

# Local save (round-trippable)
local = "./qwen-1.5b-local"
tok.save_pretrained(local)
model.save_pretrained(local, safe_serialization=True)

# Round-trip
tok2 = AutoTokenizer.from_pretrained(local)
model2 = AutoModelForCausalLM.from_pretrained(local)

# Push to Hub (need a write token; private=True for private repo)
# tok.push_to_hub("yourname/qwen-mirror", private=True)
# model.push_to_hub("yourname/qwen-mirror", private=True)

Explicit repo creation + selective upload·python

from huggingface_hub import HfApi, create_repo

repo_id = "yourname/qwen-mirror"

# Idempotent
create_repo(repo_id, repo_type="model", private=True, exist_ok=True)

api = HfApi()
api.upload_folder(
    folder_path="./qwen-1.5b-local",
    repo_id=repo_id,
    repo_type="model",
    commit_message="Initial mirror",
    ignore_patterns=["*.bin"],  # we have safetensors; skip duplicates
)

Symmetry: save_pretrained / from_pretrained

push_to_hub: the underrated half

Code

External links

Exercise

Progress

Comments 0