[release/2.8][pt2e] Pytorch 2.8 rc3 int8 accuracy is much lower than 2.7.1

[mengfei25]

🐛 Describe the bug

Tested PT2E with pytorch 2.8 rc3, mobilenet_v3_large int8 accuracy is much lower than 2.7.1.

pytorch	model	dtype	acc@1	acc@5
v2.8.0-rc3	mobilenet_v3_large	int8:	67.378	88
v2.7.1	mobilenet_v3_large	int8:	69.236	88.944

Reproducer

# setup python env
pip install --pre torch==2.8.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/xpu

git clone -b main https://github.com/chuanqi129/inductor-tools pt2e-accuracy
python pt2e-accuracy/scripts/modelbench/quant/inductor_quant_acc.py \
        --model-list mobilenet_v3_large  --device xpu --dataset_dir /path/to/imagenet

Versions

Device: PVC 1100
Driver: Rolling 25.13.33276
IGC: 2.10.11-1133~22.04
L0: 1.21.1.0-1123~22.04
Pytorch: v2.8.0-rc3

[jenniew]

@mengfei25 what imagenet dataset did you use? imagenet validation dataset? what format? Can you point out a link that I can try?

[mengfei25]

@jenniew imagenet

          wget -O valprep.sh https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh
          wget -q https://image-net.org/data/ILSVRC/2012/ILSVRC2012_img_val.tar
          tar -xf ILSVRC2012_img_val.tar
          bash valprep.sh

[jenniew]

On my environment, the test accuracy with pytorch 2.8 is better than pytorch 2.7.

Because there may be some random influence, if we don't set seed, the result varies each time. So I set seed in the testing script.

seed =1:

pytorch	model	dtype	acc@1	acc@5
v2.8.0	mobilenet_v3_large	int8:	67.762	88.150
v2.7.1	mobilenet_v3_large	int8:	67.72	88.106

seed = 1234:

pytorch	model	dtype	acc@1	acc@5
v2.8.0	mobilenet_v3_large	int8:	67.562	88.034
v2.7.1	mobilenet_v3_large	int8:	67.554	88.034

I also test without seed for 10 times, and the average accuracy with pytorch 2.8 is also better than python 2.7

pytorch	model	dtype	avg_acc@1	avg_acc@5
v2.8.0	mobilenet_v3_large	int8:	67.6436	88.0972
v2.7.1	mobilenet_v3_large	int8:	67.5448	88.0174

setup python env

2.8.0:

pip install --pre torch==2.8.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/xpu

2.7.1:

pip install --pre torch==2.7.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/xpu

Env:

Device: PVC 1550
Version: 25.18.33578.38-1146~22.04
LIBIGC: 2.11.27-1146~22.04