Accelerate Inference of Sparse Transformer Models with OpenVINO™ and 4th Gen Intel® Xeon® Scalable Processors

This tutorial is also available as a Jupyter notebook that can be cloned directly from GitHub. See the installation guide for instructions to run this tutorial locally on Windows, Linux or macOS. To run without installing anything, click the “Open in Colab” button.

This tutorial demonstrates how to improve performance of sparse Transformer models with OpenVINO on 4th Gen Intel® Xeon® Scalable processors.

The tutorial downloads a BERT-base model which has been quantized, sparsified, and tuned for SST2 datasets using Optimum-Intel. It demonstrates the inference performance advantage on 4th Gen Intel® Xeon® Scalable Processors by running it with Sparse Weight Decompression, a runtime option that seizes model sparsity for efficiency. The notebook consists of the following steps:

• Install prerequisites

• Download and quantize sparse public BERT model, using the OpenVINO integration with Hugging Face Optimum.

• Compare sparse 8-bit vs. dense 8-bit inference performance.

Table of contents:

• Prerequisites

• Imports

  • Download, quantize and sparsify the model, using Hugging Face Optimum API

• Benchmark quantized dense inference performance

• Benchmark quantized sparse inference performance

• When this might be helpful

Prerequisites ⇑

!pip install -q "openvino-dev>=2023.0.0"
!pip install -q "git+https://github.com/huggingface/optimum-intel.git" datasets onnx onnxruntime

Imports ⇑

import shutil
from pathlib import Path

from optimum.intel.openvino import OVModelForSequenceClassification
from transformers import AutoTokenizer, pipeline
from huggingface_hub import hf_hub_download

2023-08-15 22:55:04.775263: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable TF_ENABLE_ONEDNN_OPTS=0.
2023-08-15 22:55:04.809127: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
2023-08-15 22:55:05.351203: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT

INFO:nncf:NNCF initialized successfully. Supported frameworks detected: torch, tensorflow, onnx, openvino

No CUDA runtime is found, using CUDA_HOME='/usr/local/cuda'

Download, quantize and sparsify the model, using Hugging Face Optimum API. ⇑

The first step is to download a quantized sparse transformers which has been translated to OpenVINO IR. Then, it will be put through a classification as a simple validation of a working downloaded model. To find out how the model is being quantized and sparsified, refer to the OpenVINO/bert-base-uncased-sst2-int8-unstructured80 model card on Hugging Face.

# The following model has been quantized, sparsified using Optimum-Intel 1.7 which is enabled by OpenVINO and NNCF
# for reproducibility, refer https://huggingface.co/OpenVINO/bert-base-uncased-sst2-int8-unstructured80
model_id = "OpenVINO/bert-base-uncased-sst2-int8-unstructured80"

# The following two steps will set up the model and download them to HF Cache folder
ov_model = OVModelForSequenceClassification.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Let's take the model for a spin!
sentiment_classifier = pipeline("text-classification", model=ov_model, tokenizer=tokenizer)

text = "He's a dreadful magician."
outputs = sentiment_classifier(text)

print(outputs)

Compiling the model...
Set CACHE_DIR to /opt/home/k8sworker/.cache/huggingface/hub/models--OpenVINO--bert-base-uncased-sst2-int8-unstructured80/snapshots/dc44eb46300882463d50ee847e0f6485bad3cdad/model_cache
Xformers is not installed correctly. If you want to use memory_efficient_attention to accelerate training use the following command to install Xformers
pip install xformers.

[{'label': 'negative', 'score': 0.9982142448425293}]

For benchmarking, we will use OpenVINO’s benchmark application and put the IRs into a single folder.

# create a folder
quantized_sparse_dir = Path("bert_80pc_sparse_quantized_ir")
quantized_sparse_dir.mkdir(parents=True, exist_ok=True)

# following return path to specified filename in cache folder (which we've with the
ov_ir_xml_path = hf_hub_download(repo_id=model_id, filename="openvino_model.xml")
ov_ir_bin_path = hf_hub_download(repo_id=model_id, filename="openvino_model.bin")

# copy IRs to the folder
shutil.copy(ov_ir_xml_path, quantized_sparse_dir)
shutil.copy(ov_ir_bin_path, quantized_sparse_dir)

'bert_80pc_sparse_quantized_ir/openvino_model.bin'

Benchmark quantized dense inference performance ⇑

Benchmark dense inference performance using parallel execution on four CPU cores to simulate a small instance in the cloud infrastructure. Sequence length is dependent on use cases, 16 is common for conversational AI while 160 for question answering task. It is set to 64 as an example. It is recommended to tune based on your applications.

# Dump benchmarking config for dense inference
with (quantized_sparse_dir / "perf_config.json").open("w") as outfile:
    outfile.write(
        """
        {
            "CPU": {"NUM_STREAMS": 4, "INFERENCE_NUM_THREADS": 4}
        }
        """
    )

!benchmark_app -m $quantized_sparse_dir/openvino_model.xml -shape "input_ids[1,64],attention_mask[1,64],token_type_ids[1,64]" -load_config $quantized_sparse_dir/perf_config.json

huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
To disable this warning, you can either:
    - Avoid using tokenizers before the fork if possible
    - Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)
[Step 1/11] Parsing and validating input arguments
[ INFO ] Parsing input parameters
[Step 2/11] Loading OpenVINO Runtime
[ INFO ] OpenVINO:
[ INFO ] Build ................................. 2023.0.0-10926-b4452d56304-releases/2023/0
[ INFO ]
[ INFO ] Device info:
[ INFO ] CPU
[ INFO ] Build ................................. 2023.0.0-10926-b4452d56304-releases/2023/0
[ INFO ]
[ INFO ]
[Step 3/11] Setting device configuration
[ WARNING ] Performance hint was not explicitly specified in command line. Device(CPU) performance hint will be set to PerformanceMode.THROUGHPUT.
[Step 4/11] Reading model files
[ INFO ] Loading model files
[ INFO ] Read model took 74.55 ms
[ INFO ] Original model I/O parameters:
[ INFO ] Model inputs:
[ INFO ]     input_ids (node: input_ids) : i64 / [...] / [?,?]
[ INFO ]     attention_mask (node: attention_mask) : i64 / [...] / [?,?]
[ INFO ]     token_type_ids (node: token_type_ids) : i64 / [...] / [?,?]
[ INFO ] Model outputs:
[ INFO ]     logits (node: logits) : f32 / [...] / [?,2]
[Step 5/11] Resizing model to match image sizes and given batch
[ INFO ] Model batch size: 1
[ INFO ] Reshaping model: 'input_ids': [1,64], 'attention_mask': [1,64], 'token_type_ids': [1,64]
[ INFO ] Reshape model took 26.03 ms
[Step 6/11] Configuring input of the model
[ INFO ] Model inputs:
[ INFO ]     input_ids (node: input_ids) : i64 / [...] / [1,64]
[ INFO ]     attention_mask (node: attention_mask) : i64 / [...] / [1,64]
[ INFO ]     token_type_ids (node: token_type_ids) : i64 / [...] / [1,64]
[ INFO ] Model outputs:
[ INFO ]     logits (node: logits) : f32 / [...] / [1,2]
[Step 7/11] Loading the model to the device
[ INFO ] Compile model took 1231.43 ms
[Step 8/11] Querying optimal runtime parameters
[ INFO ] Model:
[ INFO ]   NETWORK_NAME: torch_jit
[ INFO ]   OPTIMAL_NUMBER_OF_INFER_REQUESTS: 4
[ INFO ]   NUM_STREAMS: 4
[ INFO ]   AFFINITY: Affinity.CORE
[ INFO ]   INFERENCE_NUM_THREADS: 4
[ INFO ]   PERF_COUNT: False
[ INFO ]   INFERENCE_PRECISION_HINT: <Type: 'float32'>
[ INFO ]   PERFORMANCE_HINT: PerformanceMode.THROUGHPUT
[ INFO ]   EXECUTION_MODE_HINT: ExecutionMode.PERFORMANCE
[ INFO ]   PERFORMANCE_HINT_NUM_REQUESTS: 0
[ INFO ]   ENABLE_CPU_PINNING: True
[ INFO ]   SCHEDULING_CORE_TYPE: SchedulingCoreType.ANY_CORE
[ INFO ]   ENABLE_HYPER_THREADING: True
[ INFO ]   EXECUTION_DEVICES: ['CPU']
[Step 9/11] Creating infer requests and preparing input tensors
[ WARNING ] No input files were given for input 'input_ids'!. This input will be filled with random values!
[ WARNING ] No input files were given for input 'attention_mask'!. This input will be filled with random values!
[ WARNING ] No input files were given for input 'token_type_ids'!. This input will be filled with random values!
[ INFO ] Fill input 'input_ids' with random values
[ INFO ] Fill input 'attention_mask' with random values
[ INFO ] Fill input 'token_type_ids' with random values
[Step 10/11] Measuring performance (Start inference asynchronously, 4 inference requests, limits: 60000 ms duration)
[ INFO ] Benchmarking in inference only mode (inputs filling are not included in measurement loop).
[ INFO ] First inference took 31.02 ms
[Step 11/11] Dumping statistics report
[ INFO ] Execution Devices:['CPU']
[ INFO ] Count:            8896 iterations
[ INFO ] Duration:         60044.19 ms
[ INFO ] Latency:
[ INFO ]    Median:        26.82 ms
[ INFO ]    Average:       26.87 ms
[ INFO ]    Min:           25.13 ms
[ INFO ]    Max:           38.04 ms
[ INFO ] Throughput:   148.16 FPS

Benchmark quantized sparse inference performance ⇑

To enable sparse weight decompression feature, users can add it to runtime config like below. CPU_SPARSE_WEIGHTS_DECOMPRESSION_RATE takes values between 0.5 and 1.0. It is a layer-level sparsity threshold for which a layer will be enabled.

# Dump benchmarking config for dense inference
# "CPU_SPARSE_WEIGHTS_DECOMPRESSION_RATE" controls minimum sparsity rate for weights to consider
# for sparse optimization at the runtime.
with (quantized_sparse_dir / "perf_config_sparse.json").open("w") as outfile:
    outfile.write(
        """
        {
            "CPU": {"NUM_STREAMS": 4, "INFERENCE_NUM_THREADS": 4, "CPU_SPARSE_WEIGHTS_DECOMPRESSION_RATE": 0.75}
        }
        """
    )

!benchmark_app -m $quantized_sparse_dir/openvino_model.xml -shape "input_ids[1,64],attention_mask[1,64],token_type_ids[1,64]" -load_config $quantized_sparse_dir/perf_config_sparse.json

huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
To disable this warning, you can either:
    - Avoid using tokenizers before the fork if possible
    - Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)
[Step 1/11] Parsing and validating input arguments
[ INFO ] Parsing input parameters
[Step 2/11] Loading OpenVINO Runtime
[ INFO ] OpenVINO:
[ INFO ] Build ................................. 2023.0.0-10926-b4452d56304-releases/2023/0
[ INFO ]
[ INFO ] Device info:
[ INFO ] CPU
[ INFO ] Build ................................. 2023.0.0-10926-b4452d56304-releases/2023/0
[ INFO ]
[ INFO ]
[Step 3/11] Setting device configuration
[ WARNING ] Performance hint was not explicitly specified in command line. Device(CPU) performance hint will be set to PerformanceMode.THROUGHPUT.
[Step 4/11] Reading model files
[ INFO ] Loading model files
[ INFO ] Read model took 63.96 ms
[ INFO ] Original model I/O parameters:
[ INFO ] Model inputs:
[ INFO ]     input_ids (node: input_ids) : i64 / [...] / [?,?]
[ INFO ]     attention_mask (node: attention_mask) : i64 / [...] / [?,?]
[ INFO ]     token_type_ids (node: token_type_ids) : i64 / [...] / [?,?]
[ INFO ] Model outputs:
[ INFO ]     logits (node: logits) : f32 / [...] / [?,2]
[Step 5/11] Resizing model to match image sizes and given batch
[ INFO ] Model batch size: 1
[ INFO ] Reshaping model: 'input_ids': [1,64], 'attention_mask': [1,64], 'token_type_ids': [1,64]
[ INFO ] Reshape model took 26.17 ms
[Step 6/11] Configuring input of the model
[ INFO ] Model inputs:
[ INFO ]     input_ids (node: input_ids) : i64 / [...] / [1,64]
[ INFO ]     attention_mask (node: attention_mask) : i64 / [...] / [1,64]
[ INFO ]     token_type_ids (node: token_type_ids) : i64 / [...] / [1,64]
[ INFO ] Model outputs:
[ INFO ]     logits (node: logits) : f32 / [...] / [1,2]
[Step 7/11] Loading the model to the device
[ INFO ] Compile model took 1252.94 ms
[Step 8/11] Querying optimal runtime parameters
[ INFO ] Model:
[ INFO ]   NETWORK_NAME: torch_jit
[ INFO ]   OPTIMAL_NUMBER_OF_INFER_REQUESTS: 4
[ INFO ]   NUM_STREAMS: 4
[ INFO ]   AFFINITY: Affinity.CORE
[ INFO ]   INFERENCE_NUM_THREADS: 4
[ INFO ]   PERF_COUNT: False
[ INFO ]   INFERENCE_PRECISION_HINT: <Type: 'float32'>
[ INFO ]   PERFORMANCE_HINT: PerformanceMode.THROUGHPUT
[ INFO ]   EXECUTION_MODE_HINT: ExecutionMode.PERFORMANCE
[ INFO ]   PERFORMANCE_HINT_NUM_REQUESTS: 0
[ INFO ]   ENABLE_CPU_PINNING: True
[ INFO ]   SCHEDULING_CORE_TYPE: SchedulingCoreType.ANY_CORE
[ INFO ]   ENABLE_HYPER_THREADING: True
[ INFO ]   EXECUTION_DEVICES: ['CPU']
[Step 9/11] Creating infer requests and preparing input tensors
[ WARNING ] No input files were given for input 'input_ids'!. This input will be filled with random values!
[ WARNING ] No input files were given for input 'attention_mask'!. This input will be filled with random values!
[ WARNING ] No input files were given for input 'token_type_ids'!. This input will be filled with random values!
[ INFO ] Fill input 'input_ids' with random values
[ INFO ] Fill input 'attention_mask' with random values
[ INFO ] Fill input 'token_type_ids' with random values
[Step 10/11] Measuring performance (Start inference asynchronously, 4 inference requests, limits: 60000 ms duration)
[ INFO ] Benchmarking in inference only mode (inputs filling are not included in measurement loop).
[ INFO ] First inference took 30.12 ms
[Step 11/11] Dumping statistics report
[ INFO ] Execution Devices:['CPU']
[ INFO ] Count:            8840 iterations
[ INFO ] Duration:         60036.27 ms
[ INFO ] Latency:
[ INFO ]    Median:        26.83 ms
[ INFO ]    Average:       26.88 ms
[ INFO ]    Min:           26.11 ms
[ INFO ]    Max:           40.45 ms
[ INFO ] Throughput:   147.24 FPS

When this might be helpful ⇑

This feature can improve inference performance for models with sparse weights in the scenarios when the model is deployed to handle multiple requests in parallel asynchronously. It is especially helpful with a small sequence length, for example, 32 and lower.

For more details about asynchronous inference with OpenVINO, refer to the following documentation:

• Deployment Optimization Guide

• Inference Request API