transformers/examples/research_projects/quantization-qdqbert
Arthur 673440d073
update ruff version (#30932)
* update ruff version

* fix research projects

* Empty

* Fix errors

---------

Co-authored-by: Lysandre <lysandre@huggingface.co>
2024-05-22 06:40:15 +02:00
..
Dockerfile Qdqbert example add benchmark script with ORT-TRT (#16592) 2022-04-12 11:13:59 -04:00
README.md Update all references to canonical models (#29001) 2024-02-16 08:16:58 +01:00
evaluate-hf-trt-qa.py update ruff version (#30932) 2024-05-22 06:40:15 +02:00
ort-infer-benchmark.py Update quality tooling for formatting (#21480) 2023-02-06 18:10:56 -05:00
quant_trainer.py update ruff version (#30932) 2024-05-22 06:40:15 +02:00
run_quant_qa.py [Docs] Fix spelling and grammar mistakes (#28825) 2024-02-02 08:45:00 +01:00
trainer_quant_qa.py Make torch xla available on GPU (#29334) 2024-03-11 14:07:16 +00:00
utils_qa.py update ruff version (#30932) 2024-05-22 06:40:15 +02:00

README.md

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:

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

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

Benchmark the INT8 QAT ONNX model inference with ONNX Runtime-TRT using dummy input

python3 ort-infer-benchmark.py

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 google-bert/bert-base-uncased \
  --int8 \
  --seed 42

Fine-tuning of FP32 model for comparison

Finetune a fp32 precision model with transformers/examples/pytorch/question-answering/:

python3 ../../pytorch/question-answering/run_qa.py \
  --model_name_or_path google-bert/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/google-bert/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/google-bert/bert-base-uncased \
  --dataset_name squad \
  --calibrator percentile \
  --percentile 99.99 \
  --max_seq_length 128 \
  --doc_stride 32 \
  --output_dir ./calib/google-bert/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/google-bert/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 google-bert/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 google-bert/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