Intel and Hugging Face Partner to Democratize Machine Learning Hardware Acceleration

Julien Simon
6 min readJun 17, 2022

This was originally posted on the Hugging Face blog.

The mission of Hugging Face is to democratize good machine learning and maximize its positive impact across industries and society. Not only do we strive to advance Transformer models, but we also work hard on simplifying their adoption.

Today, we’re excited to announce that Intel has officially joined our Hardware Partner Program. Thanks to the Optimum open-source library, Intel and Hugging Face will collaborate to build state-of-the-art hardware acceleration to train, fine-tune and predict with Transformers.

Transformer models are increasingly large and complex, which can cause production challenges for latency-sensitive applications like search or chatbots. Unfortunately, latency optimization has long been a hard problem for Machine Learning (ML) practitioners. Even with deep knowledge of the underlying framework and hardware platform, it takes a lot of trial and error to figure out which knobs and features to leverage.

Intel provides a complete foundation for accelerated AI with the Intel Xeon Scalable CPU platform and a wide range of hardware-optimized AI software tools, frameworks, and libraries. Thus, it made perfect sense for Hugging Face and Intel to join forces and collaborate on building powerful model optimization tools that let users achieve the best performance, scale, and productivity on Intel platforms.

We’re excited to work with Hugging Face to bring the latest innovations of Intel Xeon hardware and Intel AI software to the Transformers community, through open source integration and integrated developer experiences.”, says Wei Li, Intel Vice President & General Manager, AI and Analytics.

In recent months, Intel and Hugging Face collaborated on scaling Transformer workloads. We published detailed tuning guides and benchmarks on inference (part 1, part 2) and achieved single-digit millisecond latency for DistilBERT on the latest Intel Xeon Ice Lake CPUs. On the training side, we added support for Habana Gaudi accelerators, which deliver up to 40% better price-performance than GPUs.

The next logical step was to expand on this work and share it with the ML community. Enter the Optimum Intel open source library! Let’s take a deeper look at it.

Get Peak Transformers Performance with Optimum Intel

Optimum is an open-source library created by Hugging Face to simplify Transformer acceleration across a growing range of training and inference devices. Thanks to built-in optimization techniques, you can start accelerating your workloads in minutes, using ready-made scripts, or applying minimal changes to your existing code. Beginners can use Optimum out of the box with excellent results. Experts can keep tweaking for maximum performance.

Optimum Intel is part of Optimum and builds on top of the Intel Neural Compressor (INC). INC is an open-source library that delivers unified interfaces across multiple deep learning frameworks for popular network compression technologies, such as quantization, pruning, and knowledge distillation. This tool supports automatic accuracy-driven tuning strategies to help users quickly build the best quantized model.

With Optimum Intel, you can apply state-of-the-art optimization techniques to your Transformers with minimal effort. Let’s look at a complete example.

Case study: Quantizing DistilBERT with Optimum Intel

In this example, we will run post-training quantization on a DistilBERT model fine-tuned for classification. Quantization is a process that shrinks memory and compute requirements by reducing the bit width of model parameters. For example, you can often replace 32-bit floating-point parameters with 8-bit integers at the expense of a small drop in prediction accuracy.

We have already fine-tuned the original model to classify product reviews for shoes according to their star rating (from 1 to 5 stars). You can view this model and its quantized version on the Hugging Face hub. You can also test the original model in this Space.

Let’s get started! All code is available in this notebook.

As usual, the first step is to install all required libraries. It’s worth mentioning that we have to work with a CPU-only version of PyTorch for the quantization process to work correctly.

pip -q uninstall torch -y 
pip -q install torch==1.11.0+cpu --extra-index-url
pip -q install transformers datasets optimum[intel] evaluate --upgrade

Then, we prepare an evaluation dataset to assess model performance during quantization. Starting from the dataset we used to fine-tune the original model, we only keep a few thousand reviews and their labels and save them to local storage.

Next, we load the original model, its tokenizer, and the evaluation dataset.

model_name = "juliensimon/distilbert-amazon-shoe-reviews" 
model = AutoModelForSequenceClassification.from_pretrained(
model_name, num_labels=5)
tokenizer = AutoTokenizer.from_pretrained(model_name)
eval_dataset = load_from_disk("./data/amazon_shoe_reviews_test")

Next, we define an evaluation function that computes model metrics on the evaluation dataset. This allows the Optimum Intel library to compare these metrics before and after quantization. For this purpose, the Hugging Face evaluate library is very convenient!

from evaluate import evaluatordef eval_func(model): 
eval = evaluator("text-classification")
results = eval.compute(
return results["accuracy"]

We then set up the quantization job using a configuration file that we download from the Hugging Face hub. You can find details on this file in the INC documentation. Here, we go for post-training dynamic quantization with an acceptable accuracy drop of 3%.

approach: post_training_dynamic_quant
relative: 0.03

Next, we create the corresponding quantization objects with the Optimum Intel API.

from import IncOptimizer, IncQuantizer
from import IncQuantizationConfig
quantization_config = IncQuantizationConfig.from_pretrained(
inc_quantizer = IncQuantizer(quantization_config, eval_func=eval_func)

We can now launch the quantization job.

inc_optimizer = IncOptimizer(model, quantizer=inc_quantizer)
inc_model =

The log tells us that Optimum Intel has quantized 38 Linear operators.

[INFO] |*****Mixed Precision Statistics*****|
[INFO] +--------------+-----------+---------+
[INFO] | Op Type | Total | INT8 |
[INFO] +--------------+-----------+---------+
[INFO] | Linear | 38 | 38 |
[INFO] +--------------+-----------+---------+
[INFO] Pass quantize model elapsed time: 900.74 ms

Comparing the first layer of the original model (model.distilbert.transformer.layer[0]) and its quantized version (inc_model.distilbert.transformer.layer[0]), we see that Linear has indeed been replaced by DynamicQuantizedLinear, its quantized equivalent.

# Original modelTransformerBlock(
(attention): MultiHeadSelfAttention(
(dropout): Dropout(p=0.1, inplace=False)
(q_lin): Linear(in_features=768, out_features=768, bias=True)
(k_lin): Linear(in_features=768, out_features=768, bias=True)
(v_lin): Linear(in_features=768, out_features=768, bias=True)
(out_lin): Linear(in_features=768, out_features=768, bias=True)
(sa_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
(ffn): FFN(
(dropout): Dropout(p=0.1, inplace=False)
(lin1): Linear(in_features=768, out_features=3072, bias=True)
(lin2): Linear(in_features=3072, out_features=768, bias=True)
(output_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
# Quantized modelTransformerBlock(
(attention): MultiHeadSelfAttention(
(dropout): Dropout(p=0.1, inplace=False)
(q_lin): DynamicQuantizedLinear(in_features=768, out_features=768, dtype=torch.qint8, qscheme=torch.per_channel_affine)
(k_lin): DynamicQuantizedLinear(in_features=768, out_features=768, dtype=torch.qint8, qscheme=torch.per_channel_affine)
(v_lin): DynamicQuantizedLinear(in_features=768, out_features=768, dtype=torch.qint8, qscheme=torch.per_channel_affine)
(out_lin): DynamicQuantizedLinear(in_features=768, out_features=768, dtype=torch.qint8, qscheme=torch.per_channel_affine)
(sa_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
(ffn): FFN(
(dropout): Dropout(p=0.1, inplace=False)
(lin1): DynamicQuantizedLinear(in_features=768, out_features=3072, dtype=torch.qint8, qscheme=torch.per_channel_affine)
(lin2): DynamicQuantizedLinear(in_features=3072, out_features=768, dtype=torch.qint8, qscheme=torch.per_channel_affine)
(output_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)

Very well, but how does this impact accuracy and prediction time?

Before and after each quantization step, Optimum Intel runs the evaluation function on the current model. Interestingly, the accuracy of the quantized model is now a bit higher (0.6906) than the original model (0.6893). Indeed, quantization can make models more robust and less prone to over-fitting, and help increase accuracy. Likewise, we see that the quantized model predicts the evaluation set 13% faster than the original model. Not bad for a few lines of code!

[INFO] | Info Type | Baseline | Tune 1 result | Best tune result |
[INFO] | Accuracy | 0.6893 | 0.6906 | 0.6906 |
[INFO] | Duration (seconds) | 106.3149 | 92.4634 | 92.4634 |

Accuracy didn’t drop, and Optimum Intel stopped the quantization job after the first step. Had accuracy dropped more than the allowed 3%, Optimum Intel would have tried to quantize different parts of the models until it would have reached an acceptable drop, or the maximum number of trials.

Finally, we save the model and its configuration file to local storage.

model_dir = "./model_inc"

Once we’ve created a new repository on the Hugging Face hub and pushed the quantized model to it, we can load it in the usual way.

from import IncQuantizedModelForSequenceClassificationinc_model = IncQuantizedModelForSequenceClassification.from_pretrained(

We’re only getting started

In this example, we showed you how to easily quantize models post-training with Optimum Intel, and that’s just the beginning. The library supports other types of quantization as well as pruning, a technique that zeroes or removes model parameters that have little or no impact on the predicted outcome.

We are excited to partner with Intel to bring Hugging Face users peak efficiency on the latest Intel Xeon CPUs and Intel AI libraries. Please give Optimum Intel a star to get updates, and stay tuned for many upcoming features!

Many thanks to Ella Charlaix for her help on this post.