[Modeling] Load FP8 safetensors such as DeepSeek

[kylesayrs]

Purpose

• Support loading safetensors in FP8

Changes

• Add F8_E4M3 to str_to_torch_dtype mapping used by safetensors loading logic
• Add a value error Cannot load safetensors of unknown dtype {k_dtype} if string lookup fails

Testing

• Load DeepSeekV3, which has fp8 weight safetensors

from transformers import AutoModelForCausalLM

model_id = "deepseek-ai/DeepSeek-V3"
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto",
    trust_remote_code=True,
)

Reviewers

@ArthurZucker

[github-actions]

Hi 👋, thank you for opening this pull request! The pull request is converted to draft by default. When it is ready for review, please click the Ready for review button (at the bottom of the PR page).

[ArthurZucker]

LGTM, is there no requirements (like cuda specific requirements?)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[kylesayrs]

@ArthurZucker Although fp8 operations have always been someone limited and hardware dependent (gpu vs cpu), there is no hardware requirement for loading fp8 tensors outside of the required torch version

[ArthurZucker]

Will just update the branch to have the ci run!

[ArthurZucker]

Thanks @kylesayrs 🤗