U
    9%e?E                     @   s   d Z ddlmZ ddlZddlZddlmZ ddl	m
Z
 ddlmZmZmZ e rfddlZddlmZ eeZdd	 Ze re rdd
lmZ nddlmZ G dd deZG dd deZdadd Zdd Zdd Zdd Z dd Z!dddZ"dd Z#dS ) z
Integration with Deepspeed
    N)partialmethod   )dep_version_check)is_accelerate_availableis_torch_availablelogging)get_schedulerc                  C   sB   t jdd k	} | r>ztd}W dS  tjk
r<   Y dS X d S )N	deepspeedTF)	importlibutil	find_specimportlib_metadatametadataZPackageNotFoundError)Zpackage_exists_ r   b/var/www/html/Darija-Ai-API/env/lib/python3.8/site-packages/transformers/integrations/deepspeed.pyis_deepspeed_available#   s    
r   )HfDeepSpeedConfig)objectc                       s    e Zd ZdZ fddZ  ZS )r   aJ  
    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.

    c                    s(   t |  td td t | d S )NZ
accelerater	   )set_hf_deepspeed_configr   super__init__selfZconfig_file_or_dict	__class__r   r   r   I   s    zHfDeepSpeedConfig.__init__)__name__
__module____qualname____doc__r   __classcell__r   r   r   r   r   8   s   r   c                       sV   e Zd ZdZ fddZdd Zdd Zdd
dZeeddZ	dd Z
dd Z  ZS )HfTrainerDeepSpeedConfigz
    The `HfTrainerDeepSpeedConfig` object is meant to be created during `TrainingArguments` object creation and has the
    same lifespan as the latter.
    c                    s   t  | d | _g | _d S N)r   r   _dtype
mismatchesr   r   r   r   r   W   s    z!HfTrainerDeepSpeedConfig.__init__c                 C   s   | j d krtd| j S )Nz8trainer_config_process() wasn't called yet to tell dtype)r#   
ValueErrorr   r   r   r   dtype\   s    
zHfTrainerDeepSpeedConfig.dtypec                 C   s"   |  |}|d krdS |dkS d S )NFauto)	get_value)r   ds_key_longvalr   r   r   is_autoa   s    
z HfTrainerDeepSpeedConfig.is_autoNTc              
   C   s~   |  |\}}|dkrdS ||dkr4|||< dS |s<dS ||}|dk	rz||krz| jd| d| d| d|  dS )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 )Zfind_config_nodegetr$   append)r   r*   Zhf_valZhf_key
must_matchconfigZds_keyZds_valr   r   r   
fill_matchh   s    
z#HfTrainerDeepSpeedConfig.fill_matchF)r0   c                 C   s  |j |j |j }| d|jd | d|jd | d|d | d|jd | d|jd	 | d
|j|jgd | d|jd | d|j	d | 
dd | d|jd	 |js|jr|jdkrdnd}nd}|jr| jdi | jd< |j| jd d< | d|js|jo|dkd | d|dkd | d|jd | d|jpX|jd | drttj| _n| drtj| _ntj| _dS ) z
        Adjust the config with `TrainingArguments` values. This stage is run during `TrainingArguments` object
        creation.
        Ztrain_micro_batch_size_per_gpuper_device_train_batch_sizegradient_accumulation_stepstrain_batch_sizeztrain_batch_size (calculated)Zgradient_clippingmax_grad_normzoptimizer.params.lrlearning_ratezoptimizer.params.betaszadam_beta1+adam_beta2zoptimizer.params.epsadam_epsilonzoptimizer.params.weight_decayweight_decayzscheduler.params.warmup_min_lrr   zscheduler.params.warmup_max_lrZapexampN
checkpointZuse_node_local_storagezfp16.enabledz%fp16|fp16_full_eval+fp16_backend(amp)zamp.enabledzfp16+fp16_backend(apex)zamp.opt_levelfp16_opt_levelzbf16.enabledzbf16|bf16_full_eval)Z
world_sizer3   r4   r2   r6   r7   Z
adam_beta1Z
adam_beta2r8   r9   	fill_onlyZfp16Zfp16_full_evalfp16_backendZsave_on_each_noder1   r.   r<   Zbf16Zbf16_full_evalZis_truetorchZbfloat16r#   Zis_falseZfloat32Zfloat16)r   argsr5   r>   r   r   r   trainer_config_process   sD      

z/HfTrainerDeepSpeedConfig.trainer_config_processc                    s   dddg} fdd|D }t |dkrt|jdr>|jj}n*t|jdrXt|jj}ntd	| d
 d||    r dd| |   dd|   	d|d  	d|
|d t  jdkrd j}td| ddS )z
        This stage is run after we have the model and know num_training_steps.

        Now we can complete the configuration process.
        z$zero_optimization.reduce_bucket_sizez-zero_optimization.stage3_prefetch_bucket_sizez4zero_optimization.stage3_param_persistence_thresholdc                    s   g | ]}  |r|qS r   )r,   ).0xr&   r   r   
<listcomp>   s     
 zDHfTrainerDeepSpeedConfig.trainer_config_finalize.<locals>.<listcomp>r   hidden_sizehidden_sizeszThe model's config file has neither `hidden_size` nor `hidden_sizes` entry, therefore it's not possible to automatically fill out the following `auto` entries in the DeepSpeed config file: zb. You can fix that by replacing `auto` values for these keys with an integer value of your choice.g?
   z scheduler.params.total_num_stepsznum_training_steps (calculated)z!scheduler.params.warmup_num_stepsZwarmup_steps
z]Please correct the following DeepSpeed config values that mismatch TrainingArguments values:
zF
The easiest method is to set these DeepSpeed config values to 'auto'.N)lenhasattrr1   rE   maxrF   r%   r=   is_zero3r2   get_warmup_stepsr$   join)r   r@   modelnum_training_stepsZhidden_size_based_keysZhidden_size_auto_keysrE   r$   r   r&   r   trainer_config_finalize   s0    



z0HfTrainerDeepSpeedConfig.trainer_config_finalize)NT)r   r   r   r   r   r'   r,   r2   r   r=   rA   rQ   r    r   r   r   r   r!   Q   s   
:r!   c                 C   s   t | ad S r"   )weakrefref_hf_deepspeed_config_weak_ref)Zhf_deepspeed_config_objr   r   r   r      s    r   c                   C   s   d a d S r"   )rT   r   r   r   r   unset_hf_deepspeed_config   s    rU   c                   C   s$   t d k	rt  d k	rt   S dS d S )NF)rT   rL   r   r   r   r   is_deepspeed_zero3_enabled   s    
rV   c                   C   s"   t d k	rt  d k	rt  jS d S d S r"   )rT   r1   r   r   r   r   deepspeed_config  s    rW   c                    s   ddl m}m} |j}d}d|kr<|jr0td||d}n"| rNtd 	 }d|d	< d}	d
|krt||}	n4t
||r fdd}
|||
d}	nj |d}	||	fS )zY
    A convenience wrapper that deals with optimizer and lr scheduler configuration.
    r   )
DummyOptimDummySchedulerN	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)TZzero_allow_untested_optimizerZ	schedulerc                    s   t jj| j  dS )N)rZ   Znum_warmup_stepsrP   )r   r@   Zlr_scheduler_typerM   )rZ   rP   trainerr   r   _lr_scheduler_callable@  s    
z5deepspeed_optim_sched.<locals>._lr_scheduler_callable)Zlr_scheduler_callable)rP   rZ   )Zaccelerate.utilsrX   rY   r1   Z	adafactorr%   Z
is_offloadloggerinfoZcreate_optimizer
isinstanceZcreate_scheduler)r]   hf_deepspeed_configr@   rP   model_parametersrX   rY   r1   rZ   lr_schedulerr^   r   r\   r   deepspeed_optim_sched  s.    

re   Fc           
      C   s   ddl m} | j}| j}| jjjj}|||| |	|
  |rv| sTtd|d |d d\}}d}	n0d| _ttdd	 | }	t| ||||	\}}||fS )
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   )r_   zMZeRO inference only makes sense with ZeRO Stage 3 - please adjust your configrZ   rd   )NNNc                 S   s   | j S r"   )Zrequires_grad)pr   r   r   <lambda>{      z deepspeed_init.<locals>.<lambda>)Zdeepspeed.utilsr_   rO   r@   ZacceleratorstateZdeepspeed_pluginZhf_ds_configrQ   setLevelZget_process_log_levelrL   r%   Zdel_config_sub_treerZ   listfilter
parametersre   )
r]   rP   Z	inferenceZ	ds_loggerrO   r@   rb   rZ   rd   rc   r   r   r   deepspeed_initO  s.    

    rn   c                 C   sv   dd l }t| | d}t|dkrdtd|  | j|ddd\}}|d krrtd| ntd| d S )Nr   z/global_step*zAttempting to resume from T)Zload_optimizer_statesZload_lr_scheduler_statesz-[deepspeed] failed to resume from checkpoint z!Can't find a valid checkpoint at )globsortedrI   r_   r`   Zload_checkpointr%   )Zdeepspeed_engineZcheckpoint_pathro   Zdeepspeed_checkpoint_dirsZ	load_pathr   r   r   r   deepspeed_load_checkpoint  s      
rq   )F)$r   importlib.metadatar   r   importlib.utilr
   rR   	functoolsr   Zdependency_versions_checkr   utilsr   r   r   r?   optimizationr   Z
get_loggerr   r_   r   Zaccelerate.utils.deepspeedr   ZDeepSpeedConfigbuiltinsr   r!   rT   r   rU   rV   rW   re   rn   rq   r   r   r   r   <module>   s2   
 A
7