Some modules are dispatched on the CPU or the disk. Make sure you have enough GPU RAM to fit the quantized model. If you want to dispatch the model on the CPU or the disk while keeping these modules in 32-bit, you need to set load_in_8bit_fp32_cpu_offload=True and pass a custom device_map to from_pretrained. Check https://huggingface.co/docs/transformers/main/en/main_classes/quantization#offload-between-cpu-and-gpu for more details.
how to solve this error, i have 24G gpu, my transformers==4.44.2
Some modules are dispatched on the CPU or the disk. Make sure you have enough GPU RAM to fit the quantized model. If you want to dispatch the model on the CPU or the disk while keeping these modules in 32-bit, you need to set
load_in_8bit_fp32_cpu_offload=True
and pass a customdevice_map
tofrom_pretrained
. Check https://huggingface.co/docs/transformers/main/en/main_classes/quantization#offload-between-cpu-and-gpu for more details.how to solve this error, i have 24G gpu, my transformers==4.44.2