Optimization

🤗 Optimum Intel provides an openvino package that enables you to apply a variety of model quantization methods on many models hosted on the 🤗 hub using the NNCF framework.

Quantization is a technique to reduce the computational and memory costs of running inference by representing the weights and / or the activations with lower precision data types like 8-bit or 4-bit.

Weight-only quantization

Quantization can be applied on the model’s Linear, Convolutional and Embedding layers, enabling the loading of large models on memory-limited devices. For example, when applying 8-bit quantization, the resulting model will be x4 smaller than its fp32 counterpart. For 4-bit quantization, the reduction in memory could theoretically reach x8, but is closer to x6 in practice.

8-bit

For the 8-bit weight quantization you can provide quantization_config equal to OVWeightQuantizationConfig(bits=8) to load your model’s weights in 8-bit:

from optimum.intel import OVModelForCausalLM, OVWeightQuantizationConfig

model_id = "helenai/gpt2-ov"
quantization_config = OVWeightQuantizationConfig(bits=8)
model = OVModelForCausalLM.from_pretrained(model_id, quantization_config=quantization_config)

# Saves the int8 model that will be x4 smaller than its fp32 counterpart
model.save_pretrained(saving_directory)

Weights of language models inside vision-language pipelines can be quantized in a similar way:

model = OVModelForVisualCausalLM.from_pretrained(
    "llava-hf/llava-v1.6-mistral-7b-hf",
    quantization_config=quantization_config
)

4-bit

4-bit weight quantization can be achieved in a similar way:

from optimum.intel import OVModelForCausalLM, OVWeightQuantizationConfig

quantization_config = OVWeightQuantizationConfig(bits=4)
model = OVModelForCausalLM.from_pretrained(model_id, quantization_config=quantization_config)

Or for vision-language pipelines:

model = OVModelForVisualCausalLM.from_pretrained(
    "llava-hf/llava-v1.6-mistral-7b-hf",
    quantization_config=quantization_config
)

You can tune quantization parameters to achieve a better performance accuracy trade-off as follows:

quantization_config = OVWeightQuantizationConfig(
    bits=4,
    sym=False,
    ratio=0.8,
    quant_method="awq",
    dataset="wikitext2"
)

By default the quantization scheme will be asymmetric, to make it symmetric you can add sym=True.

For 4-bit quantization you can also specify the following arguments in the quantization configuration :

• The group_size parameter will define the group size to use for quantization, -1 it will results in per-column quantization.
• The ratio parameter controls the ratio between 4-bit and 8-bit quantization. If set to 0.9, it means that 90% of the layers will be quantized to int4 while 10% will be quantized to int8.

Smaller group_size and ratio values usually improve accuracy at the sacrifice of the model size and inference latency.

Quality of 4-bit weight compressed model can further be improved by employing one of the following data-dependent methods:

• AWQ which stands for Activation Aware Quantization is an algorithm that tunes model weights for more accurate 4-bit compression. It slightly improves generation quality of compressed LLMs, but requires significant additional time and memory for tuning weights on a calibration dataset. Please note that it is possible that there will be no matching patterns in the model to apply AWQ, in such case it will be skipped.
• Scale Estimation is a method that tunes quantization scales to minimize the L2 error between the original and compressed layers. Providing a dataset is required to run scale estimation. Using this method also incurs additional time and memory overhead.
• GPTQ optimizes compressed weights in a layer-wise fashion to minimize the difference between activations of a compressed and original layer.
• LoRA Correction mitigates quantization noise introduced during weight compression by leveraging low-rank adaptation.

Data-aware algorithms can be applied together or separately. For that, provide corresponding arguments to the 4-bit OVWeightQuantizationConfig together with a dataset. For example:

quantization_config = OVWeightQuantizationConfig(
    bits=4,
    sym=False,
    ratio=0.8,
    quant_method="awq",
    scale_estimation=True,
    gptq=True,
    dataset="wikitext2"
)

Note: GPTQ and LoRA Correction algorithms can’t be applied simultaneously.

Static quantization

When applying post-training static quantization, both the weights and the activations are quantized. To apply quantization on the activations, an additional calibration step is needed which consists in feeding a calibration_dataset to the network in order to estimate the quantization activations parameters.

Here is how to apply static quantization on a fine-tuned DistilBERT given your own calibration_dataset:

from transformers import AutoTokenizer
from optimum.intel import OVQuantizer, OVModelForSequenceClassification, OVConfig, OVQuantizationConfig

model_id = "distilbert-base-uncased-finetuned-sst-2-english"
model = OVModelForSequenceClassification.from_pretrained(model_id, export=True)
tokenizer = AutoTokenizer.from_pretrained(model_id)
# The directory where the quantized model will be saved
save_dir = "ptq_model"

quantizer = OVQuantizer.from_pretrained(model)

# Apply static quantization and export the resulting quantized model to OpenVINO IR format
ov_config = OVConfig(quantization_config=OVQuantizationConfig())
quantizer.quantize(ov_config=ov_config, calibration_dataset=calibration_dataset, save_directory=save_dir)
# Save the tokenizer
tokenizer.save_pretrained(save_dir)

The calibration dataset can also be created easily using your OVQuantizer:

from functools import partial

def preprocess_function(examples, tokenizer):
    return tokenizer(examples["sentence"], padding="max_length", max_length=128, truncation=True)

# Create the calibration dataset used to perform static quantization
calibration_dataset = quantizer.get_calibration_dataset(
    "glue",
    dataset_config_name="sst2",
    preprocess_function=partial(preprocess_function, tokenizer=tokenizer),
    num_samples=300,
    dataset_split="train",
)

The quantize() method applies post-training static quantization and export the resulting quantized model to the OpenVINO Intermediate Representation (IR). The resulting graph is represented with two files: an XML file describing the network topology and a binary file describing the weights. The resulting model can be run on any target Intel device.

Speech-to-text Models Quantization

The speech-to-text Whisper model can be quantized without the need for preparing a custom calibration dataset. Please see example below.

model_id = "openai/whisper-tiny"
ov_model = OVModelForSpeechSeq2Seq.from_pretrained(
    model_id,
    quantization_config=OVQuantizationConfig(
        num_samples=10,
        dataset="librispeech",
        processor=model_id,
        matmul_sq_alpha=0.95,
    )
)

With this, encoder, decoder and decoder-with-past models of the Whisper pipeline will be fully quantized, including activations.

Hybrid quantization

Traditional optimization methods like post-training 8-bit quantization do not work well for Stable Diffusion (SD) models and can lead to poor generation results. On the other hand, weight compression does not improve performance significantly when applied to Stable Diffusion models, as the size of activations is comparable to weights. The U-Net component takes up most of the overall execution time of the pipeline. Thus, optimizing just this one component can bring substantial benefits in terms of inference speed while keeping acceptable accuracy without fine-tuning. Quantizing the rest of the diffusion pipeline does not significantly improve inference performance but could potentially lead to substantial accuracy degradation. Therefore, the proposal is to apply quantization in hybrid mode for the U-Net model and weight-only quantization for the rest of the pipeline components :

• U-Net : quantization applied on both the weights and activations
• The text encoder, VAE encoder / decoder : quantization applied on the weights

The hybrid mode involves the quantization of weights in MatMul and Embedding layers, and activations of other layers, facilitating accuracy preservation post-optimization while reducing the model size.

The quantization_config is utilized to define optimization parameters for optimizing the SD pipeline. To enable hybrid quantization, specify the quantization dataset in the quantization_config. If the dataset is not defined, weight-only quantization will be applied on all components.

from optimum.intel import OVStableDiffusionPipeline, OVWeightQuantizationConfig

model = OVStableDiffusionPipeline.from_pretrained(
    model_id,
    export=True,
    quantization_config=OVWeightQuantizationConfig(bits=8, dataset="conceptual_captions"),
)

For more details, please refer to the corresponding NNCF documentation.