[BUG] Question Regarding Weights After Reloading ZeroQuant Quantized W4A8 BERT Model

Describe the bug
I loaded the weight files from the best and clean folders in the output directory, and I found that the saved weights are in FP32 format. They also don’t appear to be dequantized values. After performing torch.unique on the weights, I noticed that there are far more than 256 unique values, while after INT8 quantization, the values should only have up to 256 distinct values.

Could you clarify whether I need to perform another min-max-based quantization to extract the INT8 weights? And would this affect the accuracy reported during script execution?

To Reproduce
I used the quantization configuration from DeepSpeedExamples/compression/bert/config/ZeroQuant/ds_config_W8A8_Qgroup48_fp32.json, and the script is as follows:

TASK_NAME=mnli #mnli sst2 stsb mnli qqp rte cola mrpc qnli
export CUDA_VISIBLE_DEVICES=0
QGROUP=48
EPOCH=0
CONFIG=./config/ZeroQuant/ds_config_W8A8_Qgroup48_fp32.json
SAVE_PATH=./out/ZeroQuant/W8A8_quantization
mkdir -p ${SAVE_PATH}
MODEL=yoshitomo-matsubara/bert-base-uncased-${TASK_NAME} ## for both student and teacher
python -m torch.distributed.launch --nproc_per_node=1 \
  --use_env \
  run_glue_no_trainer.py \
  --seed 42 \
  --distill_method one_stage \
  --model_name_or_path ${MODEL} \
  --task_name $TASK_NAME \
  --max_length 128 \
  --pad_to_max_length \
  --per_device_train_batch_size 32 \
  --per_device_eval_batch_size 128 \
  --num_train_epochs ${EPOCH}\
  --eval_step 1000 \
  --deepspeed_config ${CONFIG} \
  --deepspeed \
  --save_best_model --clean_best_model \
  --gradient_accumulation_steps 1 \
  --output_dir ${SAVE_PATH} 

ds_report output

2025-02-20 06:50:26,128] [INFO] [real_accelerator.py:222:get_accelerator] Setting ds_accelerator to cuda (auto detect)
--------------------------------------------------
DeepSpeed C++/CUDA extension op report
--------------------------------------------------
NOTE: Ops not installed will be just-in-time (JIT) compiled at
      runtime if needed. Op compatibility means that your system
      meet the required dependencies to JIT install the op.
--------------------------------------------------
JIT compiled ops requires ninja
ninja .................. [OKAY]
--------------------------------------------------
op name ................ installed .. compatible
--------------------------------------------------
 [WARNING]  async_io requires the dev libaio .so object and headers but these were not found.
 [WARNING]  async_io: please install the libaio-dev package with apt
 [WARNING]  If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
async_io ............... [NO] ....... [NO]
fused_adam ............. [NO] ....... [OKAY]
cpu_adam ............... [NO] ....... [OKAY]
cpu_adagrad ............ [NO] ....... [OKAY]
cpu_lion ............... [NO] ....... [OKAY]
 [WARNING]  Please specify the CUTLASS repo directory as environment variable $CUTLASS_PATH
evoformer_attn ......... [NO] ....... [NO]
 [WARNING]  FP Quantizer is using an untested triton version (3.2.0), only 2.3.(0, 1) and 3.0.0 are known to be compatible with these kernels
fp_quantizer ........... [NO] ....... [NO]
fused_lamb ............. [NO] ....... [OKAY]
fused_lion ............. [NO] ....... [OKAY]
/home/zhaojun/anaconda3/envs/deepspeed/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `dlopen'
/home/zhaojun/anaconda3/envs/deepspeed/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `dlclose'
/home/zhaojun/anaconda3/envs/deepspeed/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `dlerror'
/home/zhaojun/anaconda3/envs/deepspeed/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `dlsym'
collect2: error: ld returned 1 exit status
gds .................... [NO] ....... [NO]
transformer_inference .. [NO] ....... [OKAY]
inference_core_ops ..... [NO] ....... [OKAY]
cutlass_ops ............ [NO] ....... [OKAY]
quantizer .............. [NO] ....... [OKAY]
ragged_device_ops ...... [NO] ....... [OKAY]
ragged_ops ............. [NO] ....... [OKAY]
random_ltd ............. [NO] ....... [OKAY]
 [WARNING]  sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.6
 [WARNING]  using untested triton version (3.2.0), only 1.0.0 is known to be compatible
sparse_attn ............ [NO] ....... [NO]
spatial_inference ...... [NO] ....... [OKAY]
transformer ............ [NO] ....... [OKAY]
stochastic_transformer . [NO] ....... [OKAY]
--------------------------------------------------
DeepSpeed general environment info:
torch install path ............... ['/home/zhaojun/anaconda3/envs/deepspeed/lib/python3.10/site-packages/torch']
torch version .................... 2.6.0+cu118
deepspeed install path ........... ['/home/zhaojun/anaconda3/envs/deepspeed/lib/python3.10/site-packages/deepspeed']
deepspeed info ................... 0.16.3, unknown, unknown
torch cuda version ............... 11.8
torch hip version ................ None
nvcc version ..................... 11.8
deepspeed wheel compiled w. ...... torch 2.5, cuda 11.8
shared memory (/dev/shm) size .... 125.79 GB

Start with config/ZeroQuant/ds_config_W8A8_Qgroup48_fp32.json and the run_glue_no_trainer.py entry point used in the reproduction. Run the provided command while tracing how the best and clean output weights are saved and loaded. Done means documenting whether the saved representation is expected, whether further quantization is required, and how it relates to reported accuracy.