MindHF stands for MindSpore + HuggingFace, representing seamless compatibility with the HuggingFace ecosystem. The name also embodies Harmonious & Fluid, symbolizing our commitment to balancing compatibility with high performance. MindHF enables you to leverage the best of both worlds: the rich HuggingFace model ecosystem and MindSpore's powerful acceleration capabilities.
Note: MindHF (formerly MindNLP) is the new name for this project. The
mindnlppackage name is still available for backward compatibility, but we recommend usingmindhfgoing forward.
MindHF provides seamless compatibility with the HuggingFace ecosystem, enabling you to run any Transformers/Diffusers models on MindSpore across all hardware platforms (GPU/Ascend/CPU) without code modifications.
You can directly use native HuggingFace libraries (transformers, diffusers, etc.) with MindSpore acceleration:
For HuggingFace Transformers:
import mindspore
import mindhf
from transformers import pipeline
chat = [
{"role": "system", "content": "You are a sassy, wise-cracking robot as imagined by Hollywood circa 1986."},
{"role": "user", "content": "Hey, can you tell me any fun things to do in New York?"}
]
pipeline = pipeline(task="text-generation", model="Qwen/Qwen3-8B", ms_dtype=mindspore.bfloat16, device_map="auto")
response = pipeline(chat, max_new_tokens=512)
print(response[0]["generated_text"][-1]["content"])For HuggingFace Diffusers:
import mindspore
import mindhf
from diffusers import DiffusionPipeline
pipeline = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", ms_dtype=mindspore.float16, device_map='cuda')
pipeline("An image of a squirrel in Picasso style").images[0]You can also use MindHF's native interface for better integration:
from mindhf.transformers import AutoTokenizer, AutoModel
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
model = AutoModel.from_pretrained("bert-base-uncased")
inputs = tokenizer("Hello world!", return_tensors='ms')
outputs = model(**inputs)Note: Due to differences in autograd and parallel execution mechanisms, any training or distributed execution code must utilize the interfaces provided by MindHF.
MindHF leverages MindSpore's powerful capabilities to deliver exceptional performance and unique features:
MindHF provides mindtorch (accessible via mindhf.core) for PyTorch-compatible interfaces, enabling seamless migration from PyTorch code while benefiting from MindSpore's acceleration on Ascend hardware:
import mindhf # Automatically enables proxy for torch APIs
import torch
from torch import nn
# All torch.xx APIs are automatically mapped to mindhf.core.xx (via mindtorch)
net = nn.Linear(10, 5)
x = torch.randn(3, 10)
out = net(x)
print(out.shape) # core.Size([3, 5])MindHF extends MindSpore with several advanced features for better model development:
- Dispatch Mechanism: Operators are automatically dispatched to the appropriate backend based on
Tensor.device, enabling seamless multi-device execution. - Meta Device Support: Perform shape inference and memory planning without actual computations, significantly speeding up model development and debugging.
- NumPy as CPU Backend: Use NumPy as a CPU backend for acceleration, providing better compatibility and performance on CPU devices.
- Heterogeneous Data Movement: Enhanced
Tensor.to()for efficient data movement across different devices (CPU/GPU/Ascend).
These features enable better support for model serialization, heterogeneous computing, and complex deployment scenarios.
You can install the official version of MindHF which is uploaded to pypi.
pip install mindhfNote: The
mindnlppackage name is still available for backward compatibility, but we recommend usingmindhfgoing forward.
You can download MindHF daily wheel from here.
To install MindHF from source, please run:
pip install git+https://github.com/mindspore-lab/mindhf.git
# or
git clone https://github.com/mindspore-lab/mindhf.git
cd mindhf
bash scripts/build_and_reinstall.sh| MindNLP version | MindSpore version | Supported Python version |
|---|---|---|
| master | daily build | >=3.7.5, <=3.9 |
| 0.1.1 | >=1.8.1, <=2.0.0 | >=3.7.5, <=3.9 |
| 0.2.x | >=2.1.0 | >=3.8, <=3.9 |
| 0.3.x | >=2.1.0, <=2.3.1 | >=3.8, <=3.9 |
| 0.4.x | >=2.2.x, <=2.5.0 | >=3.9, <=3.11 |
| 0.5.x | >=2.5.0, <=2.7.0 | >=3.10, <=3.11 |
| MindHF version | MindSpore version | Supported Python version |
|---|---|---|
| 0.6.x | >=2.7.1. | >=3.10, <=3.11 |
Since there are too many supported models, please check here
This project is released under the Apache 2.0 license.
The dynamic version is still under development, if you find any issue or have an idea on new features, please don't hesitate to contact us via Github Issues.
MindSpore NLP SIG (Natural Language Processing Special Interest Group) is the main development team of the MindHF framework. It aims to collaborate with developers from both industry and academia who are interested in research, application development, and the practical implementation of natural language processing. Our goal is to create the best NLP framework based on the domestic framework MindSpore. Additionally, we regularly hold NLP technology sharing sessions and offline events. Interested developers can join our SIG group using the QR code below.
MindSpore is an open source project that welcomes any contribution and feedback.
We wish that the toolbox and benchmark could serve the growing research
community by providing a flexible as well as standardized toolkit to re-implement existing methods
and develop their own new semantic segmentation methods.
If you find this project useful in your research, please consider citing:
@misc{mindhf2022,
title={{MindHF}: Easy-to-use and high-performance NLP and LLM framework based on MindSpore},
author={MindHF Contributors},
howpublished = {\url{https://github.com/mindspore-lab/mindnlp}},
year={2022},
note={Formerly known as MindNLP}
}