transformers/examples/research_projects/quantization-qdqbert/README.md

<!---
Copyright 2021 NVIDIA Corporation. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
-->

# Huggingface QDQBERT Quantization Example

The QDQBERT model adds fake quantization (pair of QuantizeLinear/DequantizeLinear ops) to:
 * linear layer inputs and weights
 * matmul inputs
 * residual add inputs

In this example, we use QDQBERT model to do quantization on SQuAD task, including Quantization Aware Training (QAT), Post Training Quantization (PTQ) and inferencing using TensorRT.

Required:
- [pytorch-quantization toolkit](https://github.com/NVIDIA/TensorRT/tree/master/tools/pytorch-quantization)
- [TensorRT >= 8.2](https://developer.nvidia.com/tensorrt)
- PyTorch >= 1.10.0

## Setup the environment with Dockerfile

Under the directory of `transformers/`, build the docker image:
```
docker build . -f examples/research_projects/quantization-qdqbert/Dockerfile -t bert_quantization:latest
```

Run the docker:
```
docker run --gpus all --privileged --rm -it --shm-size=1g --ulimit memlock=-1 --ulimit stack=67108864 bert_quantization:latest
```

*Note that the current NGC pytorch container (pytorch:21.07-py3) has TensorRT 8.0 which doesn't meet the requiremnt of TensorRT >= 8.2. One can either update the Dockerfile with the latest [NGC pytorch container](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch) once it supports TensorRT 8.2, or manually download and install [TensorRT >= 8.2](https://developer.nvidia.com/nvidia-tensorrt-download) in the container.*


In the container:
```
cd transformers/examples/research_projects/quantization-qdqbert/
```

## Quantization Aware Training (QAT)

Calibrate the pretrained model and finetune with quantization awared:

```
python3 run_quant_qa.py \
  --model_name_or_path bert-base-uncased \
  --dataset_name squad \
  --max_seq_length 128 \
  --doc_stride 32 \
  --output_dir calib/bert-base-uncased \
  --do_calib \
  --calibrator percentile \
  --percentile 99.99
```

```
python3 run_quant_qa.py \
  --model_name_or_path calib/bert-base-uncased \
  --dataset_name squad \
  --do_train \
  --do_eval \
  --per_device_train_batch_size 12 \
  --learning_rate 4e-5 \
  --num_train_epochs 2 \
  --max_seq_length 128 \
  --doc_stride 32 \
  --output_dir finetuned_int8/bert-base-uncased \
  --tokenizer_name bert-base-uncased \
  --save_steps 0
```

### Export QAT model to ONNX

To export the QAT model finetuned above:

```
python3 run_quant_qa.py \
  --model_name_or_path finetuned_int8/bert-base-uncased \
  --output_dir ./ \
  --save_onnx \
  --per_device_eval_batch_size 1 \
  --max_seq_length 128 \
  --doc_stride 32 \
  --dataset_name squad \
  --tokenizer_name bert-base-uncased
```

Use `--recalibrate-weights` to calibrate the weight ranges according to the quantizer axis. Use `--quant-per-tensor` for per tensor quantization (default is per channel).
Recalibrating will affect the accuracy of the model, but the change should be minimal (< 0.5 F1).

### Benchmark the INT8 QAT ONNX model inference with TensorRT using dummy input

```
trtexec --onnx=model.onnx --explicitBatch --workspace=16384 --int8 --shapes=input_ids:64x128,attention_mask:64x128,token_type_ids:64x128 --verbose
```

### Evaluate the INT8 QAT ONNX model inference with TensorRT

```
python3 evaluate-hf-trt-qa.py \
  --onnx_model_path=./model.onnx \
  --output_dir ./ \
  --per_device_eval_batch_size 64 \
  --max_seq_length 128 \
  --doc_stride 32 \
  --dataset_name squad \
  --tokenizer_name bert-base-uncased \
  --int8 \
  --seed 42
```

## Fine-tuning of FP32 model for comparison

Finetune a fp32 precision model with [transformers/examples/pytorch/question-answering/](../../pytorch/question-answering/):

```
python3 ../../pytorch/question-answering/run_qa.py \
  --model_name_or_path bert-base-uncased \
  --dataset_name squad \
  --per_device_train_batch_size 12 \
  --learning_rate 3e-5 \
  --num_train_epochs 2 \
  --max_seq_length 128 \
  --doc_stride 32 \
  --output_dir ./finetuned_fp32/bert-base-uncased \
  --save_steps 0 \
  --do_train \
  --do_eval
```

## Post Training Quantization (PTQ)

### PTQ by calibrating and evaluating the finetuned FP32 model above:

```
python3 run_quant_qa.py \
  --model_name_or_path ./finetuned_fp32/bert-base-uncased \
  --dataset_name squad \
  --calibrator percentile \
  --percentile 99.99 \
  --max_seq_length 128 \
  --doc_stride 32 \
  --output_dir ./calib/bert-base-uncased \
  --save_steps 0 \
  --do_calib \
  --do_eval
```

### Export the INT8 PTQ model to ONNX

```
python3 run_quant_qa.py \
  --model_name_or_path ./calib/bert-base-uncased \
  --output_dir ./ \
  --save_onnx \
  --per_device_eval_batch_size 1 \
  --max_seq_length 128 \
  --doc_stride 32 \
  --dataset_name squad \
  --tokenizer_name bert-base-uncased
```

### Evaluate the INT8 PTQ ONNX model inference with TensorRT

```
python3 evaluate-hf-trt-qa.py \
  --onnx_model_path=./model.onnx \
  --output_dir ./ \
  --per_device_eval_batch_size 64 \
  --max_seq_length 128 \
  --doc_stride 32 \
  --dataset_name squad \
  --tokenizer_name bert-base-uncased \
  --int8 \
  --seed 42
```

### Quantization options

Some useful options to support different implementations and optimizations. These should be specified for both calibration and finetuning.

|argument|description|
|--------|-----------|
|`--quant-per-tensor`| quantize weights with one quantization range per tensor |
|`--fuse-qkv` | use a single range (the max) for quantizing QKV weights and output activations  |
|`--clip-gelu N` | clip the output of GELU to a maximum of N when quantizing (e.g. 10) |
|`--disable-dropout` | disable dropout for consistent activation ranges |