Loading models

The Hub + from_pretrained pattern is the centre of Transformers: download (if needed), cache, and instantiate configuration, weights and preprocessors together.

Adapted for this playbook from the 🤗 Transformers documentation by Hugging Face. Official page: https://huggingface.co/docs/transformers/models. Images © Hugging Face (documentation-images) unless noted. This is not a substitute for the upstream docs — verify against the current version.

Covers Hugging Face pages: models, weightconverter, heterogeneous_configurations

Executive summary

Use Auto classes for most work. Control dtype, device map, revision and trust settings explicitly in enterprise pipelines so checkpoints are reproducible and safe.

Auto classes

from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer

model_id = "meta-llama/Llama-3.2-1B-Instruct"
config = AutoConfig.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

Pick the task head deliberately: AutoModelForSequenceClassification, AutoModelForTokenClassification, AutoModelForCausalLM, AutoModelForSeq2SeqLM, AutoModelForSpeechSeq2Seq, and so on.

Useful from_pretrained knobs

Parameter	Purpose
revision	Pin a git commit / tag for reproducibility
torch_dtype	float16, bfloat16, auto
device_map	"auto" or explicit shard map
low_cpu_mem_usage	Stream weights to reduce RAM spikes
local_files_only	Fail if not already cached (air-gap)
token / login	Access gated repos
trust_remote_code	Allow custom modelling code — review first

Dynamic weight loading

Newer Transformers releases support more flexible weight conversion and loading paths (including converting between formats). Prefer official guides when migrating checkpoints: Dynamic weight loading.

Safety checklist

• Prefer revision pins in production.
• Treat trust_remote_code=True as a security decision, not a convenience flag.
• Log model id, revision and hash in your model card / change ticket.

Key takeaways

• Auto classes resolve architecture from config.json.
• Pin revision and dtype in production.
• Never enable remote code without review.

What to do next

Customising models or Tokenizers.