o TiD@sddZddlZddlZddlmZddlmZdZddZd d Z d d Z d dZ ddZ ddZ dS)zM Functionality of swapping optimizer tensors to/from (NVMe) storage devices. N)refine_integer_value)get_accelerator:cCsD|jrt|jtkrt|j|_|jr t|jtkr t|j|_|S)N)io_sizetypestrr block_sizeargsr N/home/ubuntu/.local/lib/python3.10/site-packages/deepspeed/nvme/ds_aio_args.py refine_argss   rcsTjdurfddtjD}|Si}jD]}|t}|d||d<q|S)Ncsi|]}|jqSr )folder).0ir r r sz%_get_mapping_dict..rr)rrange multi_processfolder_to_device_mappingsplitMAPPING_DELIMITER)r dmfieldsr r r _get_mapping_dicts   rcCsd}g}dd|jD}t|dkr|dtd|d}dd|jD}d d|D}t|dkr>|d |d}|jr`d d|jD}d d|D}t|dkr`|d |d}||fS)NTcSsg|]}t|vr|qSr )rrrr r r *z,_validate_folder_mapping..rzMissing delimiter (z) in folder_to_device_mapping FcSsg|] }|tdqS)r)rrrr r r r0scSsg|] }tj|s|qSr )ospathexists)rrr r r r1sz-Invalid folders in folder_to_device_mapping: cSsg|] }t|tdqS)r)intrrrr r r r7cSsg|] }|tks|qSr )r device_count)rdev_idr r r r8r#z0Invalid device ids in folder_to_device_mapping: )rlenappendrgpu)r no_errorerror_messagesinvalid_mappings folder_listinvalid_folders device_listinvalid_device_listr r r _validate_folder_mapping's(   r0cCsd}g}|jdurt|jdkr|dd}n|jdur+t|jdkr+|dd}|jdurCtj|jsCd}|d|jdt|jdkrXt|\}}|oS|}||7}|jre|j se|dd}|st d t|d t |D]\}}t |d d |qu|S) NTrzE--folder and --folder_to_device_mapping cannot be specified together.FzIAt least one of --folder or --folder_to_device_mapping must be specified.zInvalid folder in --folder:  z,--gpu must be set to transfer with --use_gdszFound z validation errorsrz: ) rr&rr'rr r!r0use_gdsr(print enumerate)r r)r*no_mapping_errormapping_error_messagesrmsgr r r validate_args@s.     r8cCs4t}|jddtdd|jdgddd|jdtdd d d |jd d dd|jdtddd|jdtddd|jdtddd|jdd dd|jdd dd|jdd dd|jd d d!d|jd"td#d$d|jd%tdd&d|jd'd d(d|jd)d d*d|jd+d d,d|}td-||S).Nz--folderzFolder to use for I/O.)defaultrhelpz--folder_to_device_mapping+a!Specification of mapping of folder to (gpu) device id, (ignored for cpu accesses).Can be specified multiple times for multi-process runs,e.g. --folder_to_device_mapping /mnt/nvme0:0 --folder_to_device_mapping /mnt/nvme1:15 --gpumeans access /mnt/nvme0 with gpu 0 and /mnt/nvme1 with gpu 15)r9nargsr:z --io_sizeTz!Number of bytes to read or write.)rr9requiredr:z--read store_truez#Perform read I/O (default is write))actionr:z--multi_processrz3Number of parallel processes doing I/O (default 1).)rr9r:z --block_size1MzGI/O block size. Can use K, M, or G suffix (default 1M for 1 megabytes).z --queue_depth zI/O queue depth (default 32).z--single_submitzOSubmit I/O requests in singles (default is submit queue_depth amount at once.).z--sequential_requestsz}Delay I/O request submission until completion of prior requests (default is overlap I/O submission and completion requests.).z --validatez.Perform validation of I/O transfer in library.z--handlezUse AIO handle.z--loopszCount of operation repetitionsz --io_parallelzPer iop parallelismz--gpuzUse GPU memoryz --use_gdszEnable GDS AIOz--slow_bounce_bufferzSFor GPU memory transfers, measure impact of bounce buffer pinning on critical path.zargs = )argparseArgumentParser add_argumentrr" parse_argsr3)parserr r r r parse_argumentscsTrHc Cst}t|}t|sttd|jrdnd}t||_dd|jD|_ t |jt |j ks5Jtdt |j d|dt |j D]\}\}}td |d |d|d |qH|S) Nz/Successful validation of command line argumentsr(processcSsg|]\}}||fqSr r )r device_idrr r r rrz&get_validated_args..z Configuring r1z to folder mapping[z]: z <----> ) rHrr8quitr3r(r mapping_dictitems mapping_listr&r4)r peer_tagrrJrr r r get_validated_argss "rQ)__doc__rCrtest_ds_aio_utilsrdeepspeed.acceleratorrrrrr0r8rHrQr r r r s    # =