[wont merge] fp4

[wenhuach21]

wo global scale

for model_name in  "/models/opt-125m"
do
  CUDA_VISIBLE_DEVICES=$device  python3 -m auto_round \
  --model_name  $model_name \
  --device 0 \
  --tasks 'lambada_openai,hellaswag,piqa,winogrande,mmlu' \
  --bits 4 \
  --group_size 16 \
  --iters 200 \
  --format "fake" \
  --data_type "fp4_v2" \
  --act_bits 4 \
  --act_data_type "fp4_v2" \
  --output_dir "/data5/wenhuach/test_w4a4_g16" \
  --disable_deterministic_algorithms \
  2>&1 | tee -a test_fp4_g16_w4a4.txt
done

with global scale

for model_name in  "/models/opt-125m"
do
  CUDA_VISIBLE_DEVICES=$device  python3 -m auto_round \
  --model_name  $model_name \
  --device 0 \
  --tasks 'lambada_openai,hellaswag,piqa,winogrande,mmlu' \
  --bits 4 \
  --group_size 16 \
  --iters 200 \
  --format "fake" \
  --data_type "fp4_v2_with_global_scale" \
  --act_bits 4 \
  --act_data_type "fp4_v2_with_global_scale" \
  --output_dir "/data5/wenhuach/test_w4a4_g16" \
  --disable_deterministic_algorithms \
  2>&1 | tee -a test_fp4_g16_w4a4.txt
done