Weight Compression#

Weight compression is a technique for enhancing the efficiency of models, especially those with large memory requirements. This method reduces the model’s memory footprint, a crucial factor for Large Language Models (LLMs).

Weight compression is a technique for enhancing the efficiency of models, especially those with large memory requirements. This method reduces the model’s memory footprint, a crucial factor for Large Language Models (LLMs).

Unlike full model quantization, where weights and activations are quantized, weight compression in Neural Network Compression Framework (NNCF) only targets the model’s weights. This approach allows the activations to remain as floating-point numbers, preserving most of the model’s accuracy while improving its speed and reducing its size.

The reduction in size is especially noticeable with larger models, for instance the 7 billion parameter Llama 2 model can be reduced from about 25GB to 4GB using 4-bit weight compression. With smaller models (i.e. less than 1B parameters), weight compression may result in more accuracy reduction than with larger models.

LLMs and other models that require extensive memory to store the weights during inference can benefit from weight compression as it:

  • enables inference of exceptionally large models that cannot be accommodated in the device memory;

  • reduces storage and memory overhead, making models more lightweight and less resource intensive for deployment;

  • improves inference speed by reducing the latency of memory access when computing the operations with weights, for example, Linear layers. The weights are smaller and thus faster to load from memory;

  • unlike quantization, does not require sample data to calibrate the range of activation values.

Currently, NNCF provides weight quantization to 8 and 4-bit integer data types as a compression method primarily designed to optimize LLMs.

Compress Model Weights#

8-bit weight quantization method offers a balance between model size reduction and maintaining accuracy, which usually leads to significant performance improvements for Transformer-based models. Models with 8-bit compressed weights are performant on the vast majority of supported CPU and GPU platforms.

The code snippet below shows how to do 8-bit quantization of the model weights represented in OpenVINO IR using NNCF:

from nncf import compress_weights

...
model = compress_weights(model) # model is openvino.Model object

Now, the model is ready for compilation and inference. It can be also saved into a compressed format, resulting in a smaller binary file.

4-bit weight quantization method stands for an INT4-INT8 mixed-precision weight quantization, where INT4 is considered as the primary precision and INT8 is the backup one. It usually results in a smaller model size and lower inference latency, although the accuracy degradation could be higher, depending on the model.

The code snippet below shows how to do 4-bit quantization of the model weights represented in OpenVINO IR using NNCF:

from nncf import compress_weights, CompressWeightsMode

...
model = compress_weights(model, mode=CompressWeightsMode.INT4_SYM, group_size=128, ratio=0.8) # model is openvino.Model object

The table below summarizes the benefits and trade-offs for each compression type in terms of memory reduction, speed gain, and accuracy loss.

Memory Reduction

Latency Improvement

Accuracy Loss

INT8

Low

Medium

Low

INT4 Symmetric

High

High

High

INT4 Asymmetric

High

Medium

Medium

The INT4 method has several parameters that can provide different performance-accuracy trade-offs after optimization:

  • mode - there are two optimization modes: symmetric and asymmetric.

    Symmetric Compression - INT4_SYM

    INT4 Symmetric mode involves quantizing weights to an unsigned 4-bit integer symmetrically with a fixed zero point of 8. This mode is faster than the INT8, making it ideal for situations where speed and size reduction are prioritized over accuracy.

    from nncf import compress_weights
from nncf import CompressWeightsMode

compressed_model = compress_weights(model, mode=CompressWeightsMode.INT4_SYM)

    Asymmetric Compression - INT4_ASYM

    INT4 Asymmetric mode also uses an unsigned 4-bit integer but quantizes weights asymmetrically with a non-fixed zero point. This mode slightly compromises speed in favor of better accuracy compared to the symmetric mode. This mode is useful when minimal accuracy loss is crucial, but a faster performance than INT8 is still desired.

    from nncf import compress_weights
from nncf import CompressWeightsMode

compressed_model = compress_weights(model, mode=CompressWeightsMode.INT4_ASYM)

  • group_size controls the size of the group of weights that share the same quantization parameters. Shared quantization parameters help to speed up the calculation of activation values as they are dequantized and quantized between layers. However, they can reduce accuracy. The following group sizes are recommended: 128, 64, 32 (128 is default value).

    Smaller Group Size: Leads to a more accurate model but increases the model’s footprint and reduces inference speed.

    Larger Group Size: Results in faster inference and a smaller model, but might compromise accuracy.

  • ratio controls the ratio between INT4 and INT8 compressed layers in the model. Ratio is a decimal between 0 and 1. For example, 0.8 means that 80% of layers will be compressed to INT4, while the rest will be compressed to INT8 precision. The default value for ratio is 1.

    Higher Ratio (more INT4): Reduces the model size and increase inference speed but might lead to higher accuracy degradation.

    Lower Ratio (more INT8): Maintains better accuracy but results in a larger model size and potentially slower inference.

    In this example, 90% of the model’s layers are quantized to INT4 asymmetrically with a group size of 64:

    from nncf import compress_weights, CompressWeightsMode

# Example: Compressing weights with INT4_ASYM mode, group size of 64, and 90% INT4 ratio
compressed_model = compress_weights(
  model,
  mode=CompressWeightsMode.INT4_ASYM,
  group_size=64,
  ratio=0.9,
)

  • dataset - calibration dataset for data-aware weight compression. It is required for some compression options, for example, some types sensitivity_metric can use data for precision selection.

  • sensitivity_metric - controls the metric to estimate the sensitivity of compressing layers in the bit-width selection algorithm. Some of the metrics require dataset to be provided. The following types are supported:

    • nncf.SensitivityMetric.WEIGHT_QUANTIZATION_ERROR - data-free metric computed as the inverted 8-bit quantization noise. Weights with highest value of this metric can be accurately quantized channel-wise to 8-bit. The idea is to leave these weights in 8 bit, and quantize the rest of layers to 4-bit group-wise. Since group-wise is more accurate than per-channel, accuracy should not degrade.

    • nncf.SensitivityMetric.HESSIAN_INPUT_ACTIVATION - requires dataset. The average Hessian trace of weights with respect to the layer-wise quantization error multiplied by L2 norm of 8-bit quantization noise.

    • nncf.SensitivityMetric.MEAN_ACTIVATION_VARIANCE - requires dataset. The mean variance of the layers’ inputs multiplied by inverted 8-bit quantization noise.

    • nncf.SensitivityMetric.MAX_ACTIVATION_VARIANCE - requires dataset. The maximum variance of the layers’ inputs multiplied by inverted 8-bit quantization noise.

    • nncf.SensitivityMetric.MEAN_ACTIVATION_MAGNITUDE - requires dataset. The mean magnitude of the layers’ inputs multiplied by inverted 8-bit quantization noise.

  • all_layers - boolean parameter that enables INT4 weight quantization of all Fully-Connected and Embedding layers, including the first and last layers in the model.

  • awq - boolean parameter that enables the AWQ method for more accurate INT4 weight quantization. Especially helpful when the weights of all the layers are quantized to 4 bits. The method can sometimes result in reduced accuracy when used with Dynamic Quantization of activations. Requires dataset.

For data-aware weight compression refer to the following example.

The example below shows data-free 4-bit weight quantization applied on top of OpenVINO IR. Before trying the example, make sure Optimum Intel is installed in your environment by running the following command:

pip install optimum[openvino,nncf]

The first example loads a pre-trained Hugging Face model using the Optimum Intel API, compresses it to INT4 using NNCF, and then executes inference with a text phrase.

If the model comes from Hugging Face and is supported by Optimum, it can be easier to use the Optimum Intel API to perform weight compression. The compression type is specified when the model is loaded using the load_in_8bit=True or load_in_4bit=True parameter. The second example uses the Weight Compression API from Optimum Intel instead of NNCF to compress the model to INT8.

from nncf import compress_weights, CompressWeightsMode
from optimum.intel.openvino import OVModelForCausalLM
from transformers import AutoTokenizer, pipeline

# Load model from Hugging Face
model_id = "HuggingFaceH4/zephyr-7b-beta"
model = OVModelForCausalLM.from_pretrained(model_id, export=True)

# Compress to INT4 Symmetric
model.model = compress_weights(model.model,  mode=CompressWeightsMode.INT4_SYM)

# Inference
tokenizer = AutoTokenizer.from_pretrained(model_id)
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
phrase = "The weather is"
results = pipe(phrase)
print(results)

from optimum.intel.openvino import OVModelForCausalLM
from transformers import AutoTokenizer, pipeline

# Load and compress model from Hugging Face
model_id = "HuggingFaceH4/zephyr-7b-beta"
model = OVModelForCausalLM.from_pretrained(model_id, export=True, load_in_8bit=True)

# Inference
tokenizer = AutoTokenizer.from_pretrained(model_id)
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
phrase = "The weather is"
results = pipe(phrase)
print(results)

Exporting and Loading Compressed Models#

Once a model has been compressed with NNCF or Optimum Intel, it can be saved and exported to use in a future session or in a deployment environment. The compression process takes a while, so it is preferable to compress the model once, save it, and then load the compressed model later for faster time to first inference.

# Save compressed model for faster loading later
model.save_pretrained("zephyr-7b-beta-int4-sym-ov")
tokenizer.save_pretrained("zephyr-7b-beta-int4-sym-ov")

# Load a saved model
model = OVModelForCausalLM.from_pretrained("zephyr-7b-beta-int4-sym-ov")
tokenizer = AutoTokenizer.from_pretrained("zephyr-7b-beta-int4-sym-ov")

GPTQ Models#

OpenVINO also supports 4-bit models from Hugging Face Transformers library optimized with GPTQ. In this case, there is no need for an additional model optimization step because model conversion will automatically preserve the INT4 optimization results, allowing model inference to benefit from it.

A compression example using a GPTQ model is shown below. Make sure to install GPTQ dependencies by running the following command:

pip install optimum[openvino] auto-gptq

from optimum.intel.openvino import OVModelForCausalLM
from transformers import AutoTokenizer, pipeline

# Load model from Hugging Face already optimized with GPTQ
model_id = "TheBloke/Llama-2-7B-Chat-GPTQ"
model = OVModelForCausalLM.from_pretrained(model_id, export=True)

# Inference
tokenizer = AutoTokenizer.from_pretrained(model_id)
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
phrase = "The weather is"
results = pipe(phrase)
print(results)

An example of a model that has been optimized using GPTQ.

Compression Metrics Examples#

The table below shows examples of text-generation Language Models with different optimization settings in a data-free setup, where no dataset is used at the optimization step. The Perplexity metric is a measurement of response accuracy, where a higher complexity score indicates a lower accuracy. It is measured on the Lambada OpenAI dataset.

Model

Optimization

Perplexity*

Model Size (Gb)

databricks/dolly-v2-3b

FP32

5.01

10.3

databricks/dolly-v2-3b

INT8

5.07

2.6

databricks/dolly-v2-3b

INT4_ASYM,group_size=32,ratio=0.5

5.28

2.2

facebook/opt-6.7b

FP32

4.25

24.8

facebook/opt-6.7b

INT8

4.27

6.2

facebook/opt-6.7b

INT4_ASYM,group_size=64,ratio=0.8

4.32

4.1

meta-llama/Llama-2-7b-chat-hf

FP32

3.28

25.1

meta-llama/Llama-2-7b-chat-hf

INT8

3.29

6.3

meta-llama/Llama-2-7b-chat-hf

INT4_ASYM,group_size=128,ratio=0.8

3.41

4.0

togethercomputer/RedPajama-INCITE-7B-Instruct

FP32

4.15

25.6

togethercomputer/RedPajama-INCITE-7B-Instruct

INT8

4.17

6.4

togethercomputer/RedPajama-INCITE-7B-Instruct

INT4_ASYM,group_size=128,ratio=1.0

4.17

3.6

meta-llama/Llama-2-13b-chat-hf

FP32

2.92

48.5

meta-llama/Llama-2-13b-chat-hf

INT8

2.91

12.1

meta-llama/Llama-2-13b-chat-hf

INT4_SYM,group_size=64,ratio=0.8

2.98

8.0

The following table shows accuracy metric in a data-aware 4-bit weight quantization setup measured on the Wikitext dataset.

Model

Optimization

Word perplexity*

Model Size (Gb)

meta-llama/llama-7b-chat-hf

FP32

11.57

12.61

meta-llama/llama-7b-chat-hf

INT4_SYM,group_size=128,ratio=1.0,awq=True

12.34

2.6

stabilityai_stablelm-3b-4e1t

FP32

10.17

10.41

stabilityai_stablelm-3b-4e1t

INT4_SYM,group_size=64,ratio=1.0,awq=True

10.89

2.6

HuggingFaceH4/zephyr-7b-beta

FP32

9.82

13.99

HuggingFaceH4/zephyr-7b-beta

INT4_SYM,group_size=128,ratio=1.0

10.32

2.6

*Perplexity metric in both tables was measured without the Dynamic Quantization feature enabled in the OpenVINO runtime.

Auto-tuning of Weight Compression Parameters#

To find the optimal weight compression parameters for a particular model, refer to the example , where weight compression parameters are being searched from the subset of values. To speed up the search, a self-designed validation pipeline called WhoWhatBench is used. The pipeline can quickly evaluate the changes in the accuracy of the optimized model compared to the baseline.

Additional Resources#