o h?E@sdZddlmZddlZddlZddlmZddl m Z ddl m Z m Z mZe r3ddlZddlmZeeZdd Ze rIerIdd lmZndd lmZGd d d eZGdddeZdaddZddZddZddZ ddZ!dddZ"ddZ#dS) z Integration with Deepspeed N) partialmethod)dep_version_check)is_accelerate_availableis_torch_availablelogging) get_schedulercCs@tjddu}|rztd}WdStjyYdSwdS)N deepspeedTF) importlibutil find_specimportlib_metadatametadataPackageNotFoundError)package_exists_rY/var/www/html/ai/venv/lib/python3.10/site-packages/transformers/integrations/deepspeed.pyis_deepspeed_available#s r)HfDeepSpeedConfig)objectcs eZdZdZfddZZS)raJ This object contains a DeepSpeed configuration dictionary and can be quickly queried for things like zero stage. A `weakref` of this object is stored in the module's globals to be able to access the config from areas where things like the Trainer object is not available (e.g. `from_pretrained` and `_get_resized_embeddings`). Therefore it's important that this object remains alive while the program is still running. [`Trainer`] uses the `HfTrainerDeepSpeedConfig` subclass instead. That subclass has logic to sync the configuration with values of [`TrainingArguments`] by replacing special placeholder values: `"auto"`. Without this special logic the DeepSpeed configuration is not modified in any way. Args: config_file_or_dict (`Union[str, Dict]`): path to DeepSpeed config file or dict. cs(t|tdtdt|dS)N accelerater )set_hf_deepspeed_configrsuper__init__selfconfig_file_or_dict __class__rrrIszHfDeepSpeedConfig.__init__)__name__ __module__ __qualname____doc__r __classcell__rrrrr8srcsVeZdZdZfddZddZddZdd d Zeed d Z ddZ ddZ Z S)HfTrainerDeepSpeedConfigz The `HfTrainerDeepSpeedConfig` object is meant to be created during `TrainingArguments` object creation and has the same lifespan as the latter. cst|d|_g|_dSN)rr_dtype mismatchesrrrrrWs  z!HfTrainerDeepSpeedConfig.__init__cCs|jdur td|jS)Nz8trainer_config_process() wasn't called yet to tell dtype)r' ValueErrorrrrrdtype\s zHfTrainerDeepSpeedConfig.dtypecCs||}|dur dS|dkS)NFauto) get_value)r ds_key_longvalrrris_autoas z HfTrainerDeepSpeedConfig.is_autoNTc Cs||\}}|dur dS||dkr|||<dS|sdS||}|dur?||krA|jd|d|d|d|dSdSdS)a A utility method that massages the config file and can optionally verify that the values match. 1. Replace "auto" values with `TrainingArguments` value. 2. If it wasn't "auto" and `must_match` is true, then check that DS config matches Trainer config values and if mismatched add the entry to `self.mismatched` - will assert during `trainer_config_finalize` for one or more mismatches. Nr,z- ds =z vs hf )find_config_nodegetr(append)rr.hf_valhf_key must_matchconfigds_keyds_valrrr fill_matchhs  (z#HfTrainerDeepSpeedConfig.fill_matchF)r7cCs|j|j|j}|d|jd|d|jd|d|d|d|jd|d|jd |d |j|jgd |d |jd |d|j d| dd|d|jd |j s_|j ri|j dkrfdnd}nd}|jr|jdi|jd<|j|jdd<|d|j s|j o|dkd|d|dkd|d|jd|d|jp|jd|drtj|_dS|drtj|_dStj|_dS) z Adjust the config with `TrainingArguments` values. This stage is run during `TrainingArguments` object creation. train_micro_batch_size_per_gpuper_device_train_batch_sizegradient_accumulation_stepstrain_batch_sizeztrain_batch_size (calculated)gradient_clipping max_grad_normzoptimizer.params.lr learning_ratezoptimizer.params.betaszadam_beta1+adam_beta2zoptimizer.params.eps adam_epsilonzoptimizer.params.weight_decay weight_decayzscheduler.params.warmup_min_lrrzscheduler.params.warmup_max_lrapexampN checkpointuse_node_local_storagez fp16.enabledz%fp16|fp16_full_eval+fp16_backend(amp)z amp.enabledzfp16+fp16_backend(apex)z amp.opt_levelfp16_opt_levelz bf16.enabledzbf16|bf16_full_eval) world_sizer=r>r;rArB adam_beta1 adam_beta2rCrD fill_onlyfp16fp16_full_eval fp16_backendsave_on_each_noder8r3rIbf16bf16_full_evalis_truetorchbfloat16r'is_falsefloat32float16)rargsr?rPrrrtrainer_config_processs@       z/HfTrainerDeepSpeedConfig.trainer_config_processcsgd}fdd|D}t|dkrQt|jdr|jj}nt|jdr+t|jj}ntd|dd ||rQd d ||d d | d|d d| |dtj dkrwd j }td|ddS)z This stage is run after we have the model and know num_training_steps. Now we can complete the configuration process. )$zero_optimization.reduce_bucket_size-zero_optimization.stage3_prefetch_bucket_size4zero_optimization.stage3_param_persistence_thresholdcsg|] }|r|qSr)r0).0xr*rr szDHfTrainerDeepSpeedConfig.trainer_config_finalize..r hidden_size hidden_sizeszThe model's config file has neither `hidden_size` nor `hidden_sizes` entry, therefore it's not possible to automatically fill out the following `auto` entries in the DeepSpeed config file: zb. You can fix that by replacing `auto` values for these keys with an integer value of your choice.r\r]g?r^ z scheduler.params.total_num_stepsznum_training_steps (calculated)z!scheduler.params.warmup_num_steps warmup_steps z]Please correct the following DeepSpeed config values that mismatch TrainingArguments values: zF The easiest method is to set these DeepSpeed config values to 'auto'.N) lenhasattrr8rbmaxrcr)rMis_zero3r;get_warmup_stepsr(join)rrZmodelnum_training_stepshidden_size_based_keyshidden_size_auto_keysrbr(rr*rtrainer_config_finalizes4      z0HfTrainerDeepSpeedConfig.trainer_config_finalize)NT) r r!r"r#rr+r0r;rrMr[rqr$rrrrr%Qs   :r%cCst|adSr&)weakrefref_hf_deepspeed_config_weak_ref)hf_deepspeed_config_objrrrrsrcCsdadSr&)rtrrrrunset_hf_deepspeed_configsrvcCs tdurtdurtSdS)NF)rtrjrrrris_deepspeed_zero3_enableds rwcCstdur tdur tjSdSr&)rtr8rrrrdeepspeed_configsrxc sddlm}m}|j}d}d|vr|jrtd||d}n|r'td }d|d <d} d |vr=||} || fSt ||rSfd d } ||| d } || fSj |d} || fS)zY A convenience wrapper that deals with optimizer and lr scheduler configuration. r) DummyOptimDummySchedulerN optimizerz|--adafactor was passed, but also found `optimizer` configured in the DeepSpeed config. Only one optimizer can be configured.)paramszDetected ZeRO Offload and non-DeepSpeed optimizers: This combination should work as long as the custom optimizer has both CPU and GPU implementation (except LAMB)Tzero_allow_untested_optimizer schedulercstjj|jdS)N)r{num_warmup_stepsrn)rrZlr_scheduler_typerk)r{rntrainerrr_lr_scheduler_callable@s  z5deepspeed_optim_sched.._lr_scheduler_callable)lr_scheduler_callable)rnr{) accelerate.utilsryrzr8 adafactorr) is_offloadloggerinfocreate_optimizer isinstancecreate_scheduler) rhf_deepspeed_configrZrnmodel_parametersryrzr8r{ lr_schedulerrrrrdeepspeed_optim_scheds2   rFc Csddlm}|j}|j}|jjjj}||||| | |r>| s*t d| d| dd\}}d} ||fSd|_ttdd |} t||||| \}}||fS) aj Init DeepSpeed, after updating the DeepSpeed configuration with any relevant Trainer's args. If `resume_from_checkpoint` was passed then an attempt to resume from a previously saved checkpoint will be made. Args: trainer: Trainer object num_training_steps: per single gpu resume_from_checkpoint: path to a checkpoint if to resume from after normal DeepSpeedEngine load inference: launch in inference mode (no optimizer and no lr scheduler) Returns: optimizer, lr_scheduler We may use `deepspeed_init` more than once during the life of Trainer, when we do - it's a temp hack based on: https://github.com/microsoft/DeepSpeed/issues/1394#issuecomment-937405374 until Deepspeed fixes a bug where it can't resume from a checkpoint after it did some stepping https://github.com/microsoft/DeepSpeed/issues/1612 r)rzMZeRO inference only makes sense with ZeRO Stage 3 - please adjust your configr{r)NNNcSs|jSr&) requires_grad)prrr{sz deepspeed_init..)deepspeed.utilsrrmrZ acceleratorstatedeepspeed_plugin hf_ds_configrqsetLevelget_process_log_levelrjr)del_config_sub_treer{listfilter parametersr) rrn inference ds_loggerrmrZrr{rrrrrdeepspeed_initOs(      rcCstddl}t||d}t|dkr3td||j|ddd\}}|dur1td|dStd|)Nrz /global_step*zAttempting to resume from T)load_optimizer_statesload_lr_scheduler_statesz-[deepspeed] failed to resume from checkpoint z!Can't find a valid checkpoint at )globsortedrgrrload_checkpointr))deepspeed_enginecheckpoint_pathrdeepspeed_checkpoint_dirs load_pathrrrrdeepspeed_load_checkpoints  r)F)$r#importlib.metadatarr importlib.utilr rr functoolsrdependency_versions_checkrutilsrrrrU optimizationr get_loggerr rraccelerate.utils.deepspeedrDeepSpeedConfigbuiltinsrr%rtrrvrwrxrrrrrrrs4        A 7