
    Y:iV              
          S r SSKJr  SSKrSSKJr  SSKJr  SSKJrJ	r	J
r
JrJrJrJrJr  SSKJrJrJrJrJrJrJrJrJrJrJrJrJrJrJ
r
JrJrJrJrJ r J!r!JrJ"r"J#r#J$r$J%r%J&r&J'r'J(r(J)r)J*r*JrJ+r+J,r,J-r-JrJ.r.J/r/  SSK+r+SSK7  SSK0J1r1J2r2  SS	K3J4r4  SSKrSSK5r6SS
K7J8r8  SSKJr  SSK9J:r:J;r<  SSK=J>r>  SSK?r?SSK@JArA  S rB SSSSSS.rC\R                  " SS\CS9S 5       rES\R                  S\FS\FS\R                  4S jrGS\R                  S\R                  S\FS\FS\R                  4
S jrHS\R                  S\FS\R                  4S jrI\1 " S  S!\5      5       rJ  " S" S#\5      rK " S$ S%\K5      rLg)&z;
2025.10.10
2025.10.9
4.56.2
0.23.0
__UNSLOTH_VERSIONING__
    )TensorN)
functional)AnyListOptionalTupleUnionDictSetCallable)&r   BaseImageProcessorBasePairwiseJudger   DatasetEvalPredictionFFeatureExtractionMixinGeometricMixtureWrapperIterableDatasetNashMDConfigNashMDTrainerOnlineDPOTrainerOptimizerNamesr   	PeftModelPreTrainedModelPreTrainedTokenizerBaseProcessorMixinSIMPLE_CHAT_TEMPLATETrainerCallbackr	   empty_cachegenerate_model_cardget_comet_experiment_url
get_rewardis_conversationalis_peft_availableis_wandb_availablejinja2maybe_apply_chat_templatennosselective_log_softmaxtextwraptorchtruncate_rightunwrap_model_for_generation)*)	dataclassfield)Version)nullcontext)DataCollatorForSeq2SeqDataCollatorForLanguageModeling)ParallelMode)
MethodTypec                 F   ^  [         R                  " T 5      U 4S j5       nU$ )Nc                 8  > [        U S5      (       a5  [        U R                  S5      (       a  U R                  R                  5         T" U /UQ70 UD6n[        U S5      (       a5  [        U R                  S5      (       a  U R                  R                  5         U$ )Nmodelfor_trainingfor_inference)hasattrr:   r;   r<   )selfargskwargsoutputfs       K/home/james-whalen/llama.cpp/unsloth_compiled_cache/UnslothNashMDTrainer.pywrapper*prepare_for_training_mode.<locals>.wrapper0   sx     4!!gdjj.&I&IJJ##%4)$)&)4!!gdjj/&J&JJJ$$&    )	functoolswraps)rB   rD   s   ` rC   prepare_for_training_moderI   /   s%    __Q  NrF   TF)epilogue_fusionmax_autotuneshape_paddingztrace.enabledztriton.cudagraphs)dynamic	fullgraphoptionsc                 d   [         R                  " U R                  SU R                  S   5      SSS9n[         R                  " UR                  S5      SSS9n/ n[	        X#5       H  u  pVUR                  [         R                  5      n[         R                  " USUR                  S5      S9R                  S5      n[         R                  " USS9nXx-
  n	UR                  U	5        M      [         R                  " U5      nUR                  U R                  S   U R                  S   45      nU$ )N   r   )chunksdim)rT   indexrT      )r,   chunkreshapeshapeziptofloat32gather	unsqueezesqueeze	logsumexpappendconcat)
logitsrU   chunked_logitschunked_indexall_per_token_logpschunk_logitschunk_indexselected_logitslogsumexp_valuesper_token_logpss
             rC   chunked_selective_log_softmaxrm   E   s    [[FLL4D!EPQYZ[N[[r!2QaHM%(%G!#u}}5,,|2{G\G\]_G`aiijlm ??<rB)<""?3 &H 	,,':;-55v||AUV6XYrF   	input_idslogits_to_keeppad_token_idreturnc                 ~    XR                   S   :  a  [        S5      eU SS2SU* 24   nX2:H  nUR                  SS9nU$ )zr
Given prompt tensor, it returns all the left padded tokens in that sequence. so [pad, pad, pad, cat] = 3 tokens 
rW   z8logits_to_keep must be smaller than the sequence length.NrV   )rZ   
ValueErrorsum)rn   ro   rp   prompt_sectionpadding_maskpad_token_countss         rC   calculate_pad_tokens_in_promptrx   W   sX     ++STTq"2N?"223N"2L#''A'.rF   completion_input_idsleft_pad_tokens_per_promptmax_left_padc                     U R                   u  pEU R                  nX!-
  n[        R                  " XVS9R	                  S5      nXR	                  S5      :  n	X:g  n
X-  nU$ )a)  
Given that we have a sequence, [p,p,p,c,c,c,pad,pad,pad]

Where p are extra prompt tokens we got from slicing the torch tensor, c is completion tokens
and pad are pad tokens, this function would make a completion mask that would 0 out the pad
and p tokens. so in this example [0,0,0,1,1,1,0,0,0]
devicer   rW   )rZ   r~   r,   aranger_   )ry   rz   r{   rp   
batch_sizecompletion_lenr~   num_tokens_to_maskindices
shift_masknon_padding_mask
final_masks               rC    create_completion_attention_maskr   j   si     "6!;!;J!((F%Bll>9CCAFG88;;J,<.JrF   tensorpad_idc                 l    X:g  n[         R                  " USSSS9n[         R                  " U SU5      nU$ )zD
Moves all padding tokens in each sequence of a batch to the right.
rW   T)rT   
descendingstable)r,   argsortr^   )r   r   masksorted_indicespacked_tensors        rC   left_pack_paddingr      s8     D]]4Q4MNLLN;MrF   c                     ^  \ rS rSr% Sr\" SSS0S9r\\   \	S'   \" SSS	0S9r
\\   \	S
'   \" SSS0S9r\\   \	S'   SSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSS SSS!SSSSSSSSS"S"SSSSS#S$SSSSSSSS%SSSSSSSSSSSSSS%SSSSSSS&S'SSSS(SSSSSSSSSSSS)SSSSSSSSS$SSSS%SSSS*S+SSSSSSSSSSSSSSSS,S-SSSSS0 SSSS.SSS/SS0S1SS2S3S4S SSSSSSSS4U 4S5 jjrS6rU =r$ )7UnslothNashMDConfig   a  
    
Configuration class for the [`NashMDTrainer`].

Subclass of [`OnlineDPOConfig`] we can use all its arguments and add the following:

Parameters:
    mixture_coef (`float` or `list[float]`, *optional*, defaults to `0.5`):
        Logit mixture coefficient for the model and reference model. If a list of floats is provided then the
        mixture coefficient is selected for each new epoch and the last coefficient is used for the rest of the
        epochs.

    NhelpzvLLM SamplingParams)defaultmetadatavllm_sampling_paramsrQ   z8Chunk size to reduce memory usage. -1 is most efficient.unsloth_num_chunksz'Maximum sequence length to truncate to.max_seq_lengthFnorR      r      g-C6
?g{Gz?g?g+?g:0yE>g      ?g      @linear皙?passivewarningTstepsrW     iO  O1auto         
adamw_8bitlength
every_savelasti  @   i   sigmoidvllmg?colocatez0.0.0.0i@  g      n@c                   > US:  a  [        SU S35        US:  a  [        SU S35        Uc  U#S:X  a
  U$S:X  a  SnS	n#Wc$  S
SKJn  [        [	        U" 5       S-   S5      S5      nWS
::  a  [        S5      eWS:  a  [        S5      e[        TU ]  " S0 SU_SU_SU_SU_SU_SU_SU_SU_SU	_SU
_SU_SU_SU_SU_S U_S!U_S"U_S#U_S$U_S%U_S&U_S'U_S(U_S)U_S*U_S+U_S,U_S-U_S.U_S/U_S0U_S1U _S2U!_S3U"_S4U#_S5U$_S6U%_S7U&_S8U'_S9U(_S:U)_S;U*_S<U+_S=U,_S>U-_S?U._S@U/_SAU0_SBU1_SCU2_SDU3_SEU4_SFU5_SGU6_SHU7_SIU8_SJU9_SKU:_SLU;_SMU<_SNU=_SOU>_SPU?_SQW@_SRWA_SSWB_STWC_SUWD_SVWE_SWWF_SXWG_SYWH_SZWI_S[WJ_S\WK_S]WL_S^WM_S_WN_S`WO_SaWP_SbWQ_ScWR_SdWS_SeWT_SfWU_SgWV_ShWW_SiWX_SjWY_SkWZ_SlW[_SmW\_SnW]_SoW^_SpW__SqW`_SrWa_SsWb_StWc_SuWd_SvWe_SwWf_SxWg_SyWh_SzWi_S{Wj_S|Wk_S}Wl_S~Wm_SWn_SWo_SWp_SWq_SWr_SWs_SWt_SWu_SWv_SWw_SWx_SWy_SWz_SW{_SW|_SW}_SW~_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_SW_WD6  WU l        WU l	        WU l
        g )NgHz>z Unsloth: Your learning rate of `zi` is too small and less than 1e-7! Consider increasing it, otherwise gradient updates will be close to 0!rW   za` is way too larger > 1! Consider decreasing it to 1e-1, otherwise gradient updates will explode!r   r   unsloth_training_checkpointsr   r   )	cpu_countrR   r   r   zUUnsloth: Please set a positive non-zero temperature since your results will be wrong.
   zgUnsloth: Please set a positive non-zero temperature less than 10, since sampling will be quite erratic.
output_diroverwrite_output_dirdo_traindo_eval
do_predicteval_strategyprediction_loss_onlyper_device_train_batch_sizeper_device_eval_batch_sizeper_gpu_train_batch_sizeper_gpu_eval_batch_sizegradient_accumulation_stepseval_accumulation_steps
eval_delaytorch_empty_cache_stepslearning_rateweight_decay
adam_beta1
adam_beta2adam_epsilonmax_grad_normnum_train_epochs	max_stepslr_scheduler_typewarmup_ratiowarmup_steps	log_levellog_level_replicalog_on_each_nodelogging_dirlogging_strategylogging_first_steplogging_stepslogging_nan_inf_filtersave_strategy
save_stepssave_total_limitsave_safetensorssave_on_each_nodesave_only_model'restore_callback_states_from_checkpointno_cudause_cpuuse_mps_deviceseed	data_seedjit_mode_evaluse_ipexbf16fp16fp16_opt_levelhalf_precision_backendbf16_full_evalfp16_full_evaltf32
local_rankddp_backendtpu_num_corestpu_metrics_debugdebugdataloader_drop_last
eval_stepsdataloader_num_workersdataloader_prefetch_factor
past_indexrun_namedisable_tqdmremove_unused_columnslabel_namesload_best_model_at_endmetric_for_best_modelgreater_is_betterignore_data_skipfsdpfsdp_min_num_paramsfsdp_config"fsdp_transformer_layer_cls_to_wrapaccelerator_configparallelism_config	deepspeedlabel_smoothing_factoroptim
optim_args	adafactorgroup_by_lengthlength_column_name	report_toddp_find_unused_parametersddp_bucket_cap_mbddp_broadcast_buffersdataloader_pin_memorydataloader_persistent_workersskip_memory_metricsuse_legacy_prediction_looppush_to_hubresume_from_checkpointhub_model_idhub_strategy	hub_tokenhub_private_repohub_always_pushhub_revisiongradient_checkpointinggradient_checkpointing_kwargsinclude_inputs_for_metricseval_do_concat_batchesfp16_backendpush_to_hub_model_idpush_to_hub_organizationpush_to_hub_tokenmp_parametersauto_find_batch_sizefull_determinismtorchdynamo	ray_scopeddp_timeouttorch_compiletorch_compile_backendtorch_compile_modeinclude_tokens_per_secondinclude_num_input_tokens_seenneftune_noise_alphaoptim_target_modulesbatch_eval_metricseval_on_startuse_liger_kernelliger_kernel_configeval_use_gather_objectaverage_tokens_across_devicesreward_model_pathjudgemax_new_tokens
max_lengthtemperaturetop_ptop_kmin_prepetition_penaltygeneration_kwargsuse_transformers_pagedcache_implementationmissing_eos_penalty	loss_typedisable_dropoutuse_vllmvllm_model_implvllm_guided_decoding_regexvllm_gpu_memory_utilization	vllm_modevllm_server_base_urlvllm_server_hostvllm_server_portvllm_server_timeoutvllm_tensor_parallel_sizeds3_gather_for_generationmodel_init_kwargsreward_weightsdataset_num_procgpu_memory_utilization )printmultiprocessingr   minmax	MathErrorsuper__init__r   r   r   )r>   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r  r  r  r  r  r  r  r  r	  r
  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r   r!  r"  r#  r$  r%  r&  r'  r(  r)  r*  r+  r,  r-  r.  r/  r0  r1  r2  r3  r4  r5  r6  r7  r8  r9  r:  r;  r<  r=  r>  r?  r@  rA  rB  rC  rD  rE  rF  rG  rH  rI  rJ  rK  rL  rM  rN  rO  rP  rQ  rR  r   r   r   r@   r   	__class__s                                                                                                                                                                        rC   rZ  UnslothNashMDConfig.__init__   sr	   L 4)I-  YB  (C  "D1e&F}o  Vw  %x  y-7":zS?P7J M#1"3y{1}a#8"=!sttB  F  G  G 	 _	F#_	F#7_	F  _	F 	_	F
 $_	F *_	F $8_	F +F_	F *D_	F (@_	F '>_	F +F_	F '>_	F $_	F '>_	F  *!_	F" (#_	F$ $%_	F& $'_	F( ()_	F* *+_	F,  0-_	F. "/_	F0 !21_	F2 (3_	F4 (5_	F6 "7_	F8 !29_	F:  0;_	F< &=_	F>  0?_	F@ "4A_	FB *C_	FD &<E_	FF *G_	FH $I_	FJ  0K_	FL  0M_	FN !2O_	FP .Q_	FR 7^S_	FT U_	FV W_	FX ,Y_	FZ [_	F\ "]_	F^ *__	F`  a_	Fb c_	Fd e_	Ff ,g_	Fh &<i_	Fj ,k_	Fl ,m_	Fn o_	Fp $q_	Fr &s_	Ft *u_	Fv !2w_	Fx y_	Fz $8{_	F| $}_	F~ &<_	F@ *DA_	FB $C_	FD  E_	FF (G_	FH %:I_	FJ &K_	FL &<M_	FN %:O_	FP !2Q_	FR  0S_	FT U_	FV #6W_	FX &Y_	FZ 2T[_	F\ "4]_	F^ "4__	F` "a_	Fb &<c_	Fd e_	Ff $g_	Fh "i_	Fj .k_	Fl "4m_	Fn "o_	Fp *Dq_	Fr !2s_	Ft %:u_	Fv %:w_	Fx -Jy_	Fz #6{_	F| *D}_	F~ &_	F@ &<A_	FB (C_	FD (E_	FF "G_	FH  0I_	FJ .K_	FL (M_	FN &<O_	FP -JQ_	FR *DS_	FT &<U_	FV (W_	FX $8Y_	FZ (@[_	F\ !2]_	F^ *__	F` $8a_	Fb  0c_	Fd &e_	Ff "g_	Fh &i_	Fj *k_	Fl %:m_	Fn "4o_	Fp )Bq_	Fr -Js_	Ft #6u_	Fv $8w_	Fx "4y_	Fz *{_	F|  0}_	F~ #6_	F@ &<A_	FB -JC_	FD !2E_	FF G_	FH ,I_	FJ $K_	FL &M_	FN O_	FP Q_	FR S_	FT "4U_	FV !2W_	FX &<Y_	FZ $8[_	F\ #6]_	F^ "__	F` .a_	Fb  c_	Fd .e_	Ff *Dg_	Fh +Fi_	Fj "k_	Fl $8m_	Fn  0o_	Fp  0q_	Fr #6s_	Ft )Bu_	Fv )Bw_	Fx !2y_	Fz ,{_	F|  0}_	F~ &<f_	F@ %9!"4,rF   )r   r   r   )__name__
__module____qualname____firstlineno____doc__r1   r   r   r   __annotations__r   intr   rZ  __static_attributes____classcell__r[  s   @rC   r   r      sh    +012+(3-  */VW*#  &+EF&NXc]  #$&'%&#'"&&'"#"%$%""!&!27!'!$!"%) $!& $  -1!!!$%%)  $ $(-"%*!%#!%(,%*!%##' $  $!$)(-"#" "!&(,  !&#" %)&*#$#$%$( !%#GV- V-rF   r   c            "       B  ^  \ rS rSrSrSS/r               S(S\\\R                  4   S\\\R                  4   S	\\\R                  S4   S
\
\   S\
\   S\
\   S\
\\\4      S\
\\\\\4   4      S\
\\\\\4      S\
\   S\
\\/\4      S\
\\      S\\R6                  R8                  \R6                  R:                  R<                  4   S\
\\R>                  \R>                  /\R>                  4      S\
\\\R                  4      SS4 U 4S jjjr \!S 5       r"S r#S r$S r%S r&S r'S r(  SS jr) S)S\R                  S \\\\R>                  \*4   4   S!\
\+   S\R>                  4S" jjr,   S*S#\
\   S$\
\   S%\\\\   S4   4S& jjr-S'r.U =r/$ )+_UnslothNashMDTraineri  a
  
Initialize NashMDTrainer as a subclass of [`OnlineDPOConfig`].

Args:
    model (`transformers.PreTrainedModel`):
        The model to train, preferably an `AutoModelForCausalLM`.
    ref_model (`PreTrainedModelWrapper`):
        Hugging Face transformer model with a casual language modelling head. Used for implicit reward computation
        and loss. If no reference model is provided, the trainer will create a reference model with the same
        architecture as the model to be optimized.
    reward_funcs (`transformers.PreTrainedModel`):
        The reward model to score completions with, preferably an `AutoModelForSequenceClassification`.
    judge (`BasePairwiseJudge`):
        The judge to use for pairwise comparison of model completions.
    args (`NashMDConfig`):
        The NashMD config arguments to use for training.
    data_collator (`transformers.DataCollator`):
        The data collator to use for training. If None is specified, the default data collator
        (`DPODataCollatorWithPadding`) will be used which will pad the sequences to the maximum length of the
        sequences in the batch, given a dataset of paired sequences.
    train_dataset (`datasets.Dataset`):
        The dataset to use for training.
    eval_dataset (`datasets.Dataset`):
        The dataset to use for evaluation.
    processing_class ([`~transformers.PreTrainedTokenizerBase`], [`~transformers.BaseImageProcessor`], [`~transformers.FeatureExtractionMixin`] or [`~transformers.ProcessorMixin`], *optional*, defaults to `None`):
        Processing class used to process the data. If provided, will be used to automatically process the inputs
        for the model, and it will be saved along the model to make it easier to rerun an interrupted training or
        reuse the fine-tuned model.
    peft_config (`dict`):
        The peft config to use for training.
    compute_metrics (`Callable[[EvalPrediction], dict]`, *optional*):
        The function to use to compute the metrics. Must take a `EvalPrediction` and return a dictionary string to
        metric values.
    callbacks (`list[transformers.TrainerCallback]`):
        The callbacks to use for training.
    optimizers (`tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR]`):
        The optimizer and scheduler to use for training.
    preprocess_logits_for_metrics (`Callable[[torch.Tensor, torch.Tensor], torch.Tensor]`):
        The function to use to preprocess the logits before computing the metrics.

.. deprecated:: 0.22.0
    The following parameters are deprecated and will be removed in a future version:

    * `reward_model`: Use `reward_funcs` instead. For example, change `reward_model=model` to `reward_funcs=model`.
trlznash-mdNNNr:   	ref_modelreward_funcsr6  r?   data_collatortrain_dataseteval_datasetprocessing_classpeft_configcompute_metrics	callbacks
optimizerspreprocess_logits_for_metricsreward_modelrq   c                 d  > [         TU ]  UUUUUUUUU	U	U
UUUUUS9  U R                  R                  U l        / / / / / / / / / / / / S.U l        U R                  bW  [        U R                  5      S:w  a  [        S5      eU R                  S   U l        / U R
                  S'   / U R
                  S'   g g )N)r:   rk  rl  r6  r?   rm  rn  ro  rp  reward_processing_classesrq  rr  rs  rt  ru  rv  )loss/klobjective/entropy
loss/scorerewards/probabilitiesrewards/accuraciesrewards/marginslogps/chosenlogps/rejectedval/model_contain_eos_tokenval/ref_contain_eos_tokenbetamixture_coefrW   z6NashMDTrainer only supports one reward function/model.r   rewards/chosenrewards/rejected)	rY  rZ  r?   r  _mixture_coefstatsrl  lenrs   )r>   r:   rk  rl  r6  r?   rm  rn  ro  rp  rq  rr  rs  rt  ru  rv  r[  s                   rC   rZ  _UnslothNashMDTrainer.__init__3  s    * 	%''%-&6#+!*G%! 	 	
& "YY33 !#%'"$! +-)+

  (4$$%* !YZZ $ 1 1! 4D+-DJJ'(-/DJJ)* )rF   c                     [        U R                  [        5      (       aM  U R                  R                  nU[        U R                  5      :  a  U R                  U   $ U R                  S   $ U R                  $ )NrQ   )
isinstancer  liststateepochr  )r>   r  s     rC   r  "_UnslothNashMDTrainer.mixture_coefu  se    d(($//JJ$$E05D<N<N8O0O4%%e,kUYUgUghjUkk%%%rF   c           	         [        XR                  5       nUR                  US   US   U R                  S9nS S S 5        U R                  R	                  U5      nU R
                  c8  [        5       (       a&  [        U[        5      (       a  UR                  5       nO(UnO%U R                  R	                  U R
                  5      n[        R                  " 5          [        UUU R                  U R                  U R                  R                  S9nUR                  US   US   U R                  S9nS S S 5        WU4$ ! , (       d  f       GN= f! , (       d  f       WW4$ = f)Nrn   attention_mask)rn   r  generation_config)r:   rk  r  r  r~   )r.   acceleratorgenerater  unwrap_modelrk  r$   r  r   get_base_modelr,   no_gradr   r  r~   )	r>   r:   promptsunwrapped_policy_for_gen_ctxmodel_outputpolicy_model_for_gmwref_model_for_gmwmixture_modelmixture_outputs	            rC   _generate_completions+_UnslothNashMDTrainer._generate_completions}  sY   (0@0@AEa7@@!+.&'78"&"8"8 A L B  $//<<UC
 >>! !""z2F	'R'R$8$G$G$I! %9! !% 0 0 = =dnn M ]]_3*+"&"8"8!..''..M +33!+.&'78"&"8"8 4 N  ^++Y BA: _ ^++s   "D=AE=
E
E c                    US   R                   S   nUS S 2US 24   n[        XPR                  R                  U R                  R                  5      u  pV[
        R                  " US   U4SS9[
        R                  " US   U4SS9US   S.nUS S 2US 24   n[        XR                  R                  U R                  R                  5      u  p[
        R                  " US   U4SS9[
        R                  " US   U	4SS9US   S.n
Xz4$ )Nrn   rW   rV   r  rawrn   r  r  )rZ   r-   rp  eos_token_idrp   r,   cat)r>   r  r  r  context_lengthmodel_completion_idsmodel_completion_mask
model_datamixture_completion_idsmixture_completion_maskmixture_datas              rC   _process_completions*_UnslothNashMDTrainer._process_completions  s*    -33A6  ,A~,>?6D "7"7"D"DdF[F[FhFh7
3 GK$8:N#OUVW#ii1A)BDY(Z`ab5>

 "0>?0B!C:H"$9$9$F$FH]H]HjHj;
7 GK$8:P#QWXY#ii1A)BD[(\bcd5>
 ''rF   c                    [         R                  " 5          [        U R                  US   U R                  R
                  U5      u  pEn[        U R                  US   U R                  R
                  U5      u  pFnS S S 5        U R                  R                  b  [         R                  " US   U R                  R                  :H  SS9n[         R                  " US   U R                  R                  :H  SS9nWU) ==   U R                  R                  -  ss'   WU) ==   U R                  R                  -  ss'   WW4$ ! , (       d  f       N= f)Nrn   rQ   rV   )
r,   r  r"   rl  rp  rp   r?   rA  anyr  )	r>   r  r  r  _model_scoresmixture_scoresmodel_contain_eosmixture_contain_eoss	            rC   _compute_rewards&_UnslothNashMDTrainer._compute_rewards  s#   ]]_!+!!:k#:D<Q<Q<^<^`n"AQ $.!!<#<d>S>S>`>`bp$ Aq	  99((4 %		*[*ATEZEZEgEg*gmo p"'))L,EI^I^IkIk,kqs"t++,		0M0MM,//0DII4Q4QQ0^++ _s   A%E  
Ec           	         US   nU R                   R                  US   S S 2US 24   SS9nU Vs/ s H  ofR                  5       PM     nnU R                   R                  US   S S 2US 24   SS9nU Vs/ s H  ofR                  5       PM     nn[        SUS   05      (       a  U Vs/ s H	  nSUS./PM     nn[        R
                  " 5       nUR                  [        5      n	U V
s/ s H  oR                  U
S	9PM     nn
U Vs/ s H  oiR                  US	9PM     nnU Vs/ s H	  nSUS./PM     nnU Vs/ s H  oiR                  US	9PM     nnU R                  R                  U[        [        XW5      5      SS
9n[        R                  " XS   R                  S9$ s  snf s  snf s  snf s  sn
f s  snf s  snf s  snf )Nr  rn   T)skip_special_tokenspromptr   	assistant)rolecontent)messages)return_scoresr}   )rp  batch_decodestripr#   r&   Environmentfrom_stringr   renderr6  r  r[   r,   r   r~   )r>   r  r  r  r  model_data_completions
completionmixture_data_completionsenvironmenttemplatemessageprobabilitys               rC   _compute_judge$_UnslothNashMDTrainer._compute_judge  s   U#!%!6!6!C!C{#A~$67T "D "
 H^!^G]"2"2"4G]!^#'#8#8#E#E%a&89t $F $
  Jb#bIa:$4$4$6Ia #bh
344Qg&Qg:+*=>Qg # & !,,.K"../CDHHOPW8GP]s%t]szoozo&J]s"%t Rj(Qi:+*=>Qi % ( H`(G_4G_ % ( jj&&+FG ' 

 ||K;0G0N0NOO7 "_
 $c&
 Q%t((s)   F#5F()F-)F2F7%F<;Gc                   ^ U4S jnU" X5      n[         R                  " 5          U R                  c"  UR                  5          U" X5      nS S S 5        OU" U R                  U5      nS S S 5        US   S S 2TS 24   S:H  nUR	                  US5      nWR	                  US5      nXV4$ ! , (       d  f       NN= f! , (       d  f       NW= f)Nc                    > U " US   US   S9nUR                   S S 2TS-
  S24   n[        X1S   S S 2TS 24   5      nU$ )Nrn   r  )r  rW   rQ   )rd   r*   )mdatarA   rd   token_logprobsr  s        rC   compute_logprobs_for_dataJ_UnslothNashMDTrainer._compute_logprobs.<locals>.compute_logprobs_for_data  sZ    tK(>N9OPF]]1nq&82&=#=>F26;LQP^P_M_;`aN!!rF   r  r   r   )r,   r  rk  disable_adaptermasked_fill)r>   r:   r  r  r  model_logprobs_model_dataref_logprobs_model_datamodel_padding_masks      `    rC   _compute_logprobs'_UnslothNashMDTrainer._compute_logprobs  s    	" %>e$P! ]]_~~%**,.G.Z+ -, +DDNNT^*_'  ((89!^_:LMQRR$=$I$IJ\^a$b!"9"E"EFXZ]"^)CC -, _s#   B9	B(B9(
B6	2B99
Cc                 *   US-
  UR                  S5      -  n[        R                  " 5          X-
  nUR                  S5      nS S S 5        WU-  R                  S5      nU R                  U-  U-
  nUR	                  5       UW4$ ! , (       d  f       NG= f)Ng      ?rW   )rt   r,   r  r  mean)	r>   r  r  r  score	log_ratio
kl_div_logkl_div_losslosss	            rC   _compute_losses%_UnslothNashMDTrainer._compute_losses  s     s"&?&C&CA&FF ]]_1KI"q)J  !#<<AA!D yy;&.yy{E:-- _s   B
Bc                   ^  U 4S jnT R                   S   R                  U" U5      5        T R                   S   R                  U" U5      5        UR                  S5      nUR                  S5      nT R                   S   R                  U" U5      5        T R                   S   R                  U" U5      5        T R                  bH  T R                   S   R                  U" U	5      5        T R                   S   R                  U" U
5      5        T R                   S	   R                  U" U5      5        UR                  S5      * nT R                   S
   R                  U" U5      5        X-
  nT R                   S   R                  U" U5      5        US:  R	                  5       nT R                   S   R                  U" U5      5        US   S S 2US 24   T R
                  R                  :H  R                  SS9nUS   S S 2US 24   T R
                  R                  :H  R                  SS9nT R                   S   R                  U" UR	                  5       5      5        T R                   S   R                  U" UR	                  5       5      5        T R                   S   R                  T R                  5        T R                   S   R                  T R                  5        g )Nc                 r   > TR                   R                  U 5      R                  5       R                  5       $ N)r  gather_for_metricsr  item)r   r>   s    rC   gather_mean:_UnslothNashMDTrainer._log_statistics.<locals>.gather_mean7  s,    ##66v>CCEJJLLrF   r{  ry  rW   r  r  r  r  r|  rz  r~  r   r}  rn   rV   r  r  r  r  )
r  rb   rt   rl  floatrp  r  r  r  r  )r>   r  r  r  r  r  r  kl_divr  r  r  r  model_logprobs_model_data_sumref_logprobs_model_data_sumentropy_model_datamarginaccuracy	model_eosmixture_eoss   `                  rC   _log_statistics%_UnslothNashMDTrainer._log_statistics)  s   	M 	

< ''E(:;

9$$[%89 )B(E(Ea(H%&=&A&A!&D#

>"))+6S*TU

#$++K8S,TU (JJ'(//L0IJJJ)*11+n2MN 	

*+22;{3KL 8;;A>>

&'..{;M/NO /L

$%,,[-@A QJ%%'

'(//H0EF  ,Q-?@DDYDYDfDffkkpqkr	#K0NO1CDH]H]HjHjjootuov

0188Y__EV9WX

./66{;CTCTCV7WX 	

6!!$)),

>"))$*;*;<rF   inputsnum_items_in_batchc                 <   UR                  5         [        [        [        UR	                  5       5      5      5      nUS   n[        U5       VVVs/ s H*  obR                  5        VVs0 s H
  u  pxXxU   _M     snnPM,     nnnnU V	s/ s H  n	[        XR                  5      PM     nn	U V	s/ s H<  oR                  XR                  R                  R                  U R                  5      PM>     nn	U R                  U5      nU R                  U5      nUS   R                  S   n
US   US   US.nAU R!                  X5      u  pU R#                  XU5      u  pU R$                  b/  U R'                  XU
5      u  nn[(        R*                  " UU-
  5      nOSu  nnU R-                  XU
5      nU R/                  XU
5      u  nnU R1                  UUU5      u  nnnU R3                  UUUR5                  5       UUUR5                  5       UR5                  5       U
UU5
        U R6                  R8                  b;  U R:                  R<                  U R6                  R8                  -  S:X  a
  [?        5         0 nU R6                  R@                  [B        RD                  [B        RF                  4;   a  U RI                  5       US'   U R6                  RJ                  S:  a  URM                  5       nU RN                  (       a:  [P        RS                  UU RT                  5       nURW                  5         S S S 5        OU RX                  RV                  " U40 UD6  UR5                  5       U R6                  RZ                  -  $ s  snnf s  snnnf s  sn	f s  sn	f ! , (       d  f       NL= f)	Nr  prompt_input_idsrW   prompt_attention_maskr  rj  r   r   ).trainr  nextitervaluesrangeitemsr'   rp  tokenize_rowr:   configis_encoder_decoderrm  _prepare_inputsrZ   r  r  rl  r  r   r   r  r  r  r  detachr?   r   r  global_stepr   r  r   LOMOADALOMO_get_learning_raten_gpur  use_apexamp
scale_loss	optimizerbackwardr  r   )r>   r:   r  r  r   r  ikvxr  r  r  r  r  r  r  r  r  r  r  r  r  r@   scaled_losss                            rC   training_step#_UnslothNashMDTrainer.training_stepd  s'    	 d6==?345
"@Ej@QR@Q1||~6~tq1d7~6@QROUVv!+A/D/DEvVmstmshi##Azz'8'8'K'KTMbMbcmst##F+ %%f- 2399!< 23$%<=

  (,'A'A%'Q$ $(#<#<\[b#c 
 (+/+@+@[i+j(L.))L>$ABK+5(L.--jWK >B=S=STYgu=v:!#: #223LNegrseV 	%,,.#LLNMMO	
 II--9

&&)J)JJaOM99??~22N4J4JKK&*&=&=&?F?#99??Q99;D==dnn5$$& 65 %%d5f5{{}tyyDDDDM 7RVt~ 65s1   M<(M69M<	N.ANN6M<
N
model_namedataset_nametagsc                    U R                  5       (       d  g[        U R                  R                  S5      (       ac  [        R
                  R                  U R                  R                  R                  5      (       d!  U R                  R                  R                  nOSnUc  [        5       nO$[        U[        5      (       a  U1nO[        U5      n[        U R                  R                  S5      (       a  UR                  S5        S[        R                  ;   a  UR                  S5        UR                  U R                  5        [        R                   " S5      n[#        UUU R$                  UU['        5       (       a+  [(        R*                  b  [(        R*                  R,                  OS[/        5       SUS	S
S9nUR1                  [        R
                  R3                  U R4                  R6                  S5      5        g)a  
Creates a draft of a model card using the information available to the `Trainer`.

Args:
    model_name (`str` or `None`, *optional*, defaults to `None`):
        Name of the model.
    dataset_name (`str` or `None`, *optional*, defaults to `None`):
        Name of the dataset used for training.
    tags (`str`, `list[str]` or `None`, *optional*, defaults to `None`):
        Tags to be associated with the model card.
N_name_or_pathunsloth_versionunslothJOB_IDhf_jobsa          @inproceedings{munos2024nash,
            title        = {{Nash Learning from Human Feedback}},
            author       = {R{'{e}}mi Munos and Michal Valko and Daniele Calandriello and Mohammad Gheshlaghi Azar and Mark Rowland and Zhaohan Daniel Guo and Yunhao Tang and Matthieu Geist and Thomas Mesnard and C{\^{o}}me Fiegel and Andrea Michi and Marco Selvi and Sertan Girgin and Nikola Momchev and Olivier Bachem and Daniel J. Mankowitz and Doina Precup and Bilal Piot},
            year         = 2024,
            booktitle    = {Forty-first International Conference on Machine Learning, {ICML} 2024, Vienna, Austria, July 21-27, 2024},
            publisher    = {OpenReview.net},
            url          = {https://openreview.net/forum?id=Y5AmNYiyCQ}
        }zNash-MDz!Nash Learning from Human Feedbackz
2312.00886)
base_modelr  r  r  r  	wandb_url	comet_urltrainer_nametrainer_citationpaper_titlepaper_idz	README.md)is_world_process_zeror=   r:   r  r)   pathisdirr  setr  straddenvironupdate
_tag_namesr+   dedentr    r  r%   wandbrunurlr!   savejoinr?   r   )r>   r  r  r  r"  citation
model_cards          rC   create_model_card'_UnslothNashMDTrainer.create_model_card  sn   " ))++4::$$o66rww}}TZZM^M^MlMl?m?m**88JJ <5Dc""6Dt9D4::$$&788HHYrzz!HHYDOO$ ?? $  )!!**%'9';';		@Ueiimm[_.0"%;!

 	TYY%9%9;GHrF   )r  rl  r  )NNNNNNNNNNNNrj  NNr  )NNN)0r]  r^  r_  r`  ra  r1  r	   r   r(   Moduler   r   r   r   r   r   dictr-  r   r   r   r   r   r  r   tupler,   r  	Optimizerlr_schedulerLambdaLRr   rZ  propertyr  r  r  r  r  r  r  r  r   rc  r  r:  rd  re  rf  s   @rC   rh  rh    s   ,\ #J 487;@D-1'+,0CGEI &*FJ59VbhlDH'@0_bii/0@0 "))34@0 ORYY<=	@0
 )*@0 |$@0  )@0  g&> ?@@0 uWd3<.@%@AB@0 #)+=?UWeef
@0 d^@0 "(N+;T+A"BC@0 D12@0  %++//1I1I1R1RRS!@0" (0%,,9UW\WcWc9c0d'e#@0& u_bii%?@A'@0( 
)@0 @0D & &.,`(6,$ PDD2.< 9=x rvNEYYNE(,S%c8I2J-J(KNEaijmanNE	NEd %)&*,0	BISMBI smBI CcD()	BI BIrF   rh  c                   L   ^  \ rS rSrSr              SU 4S jjrSrU =r$ )UnslothNashMDTraineri  a
  
    
Initialize NashMDTrainer as a subclass of [`OnlineDPOConfig`].

Args:
    model (`transformers.PreTrainedModel`):
        The model to train, preferably an `AutoModelForCausalLM`.
    ref_model (`PreTrainedModelWrapper`):
        Hugging Face transformer model with a casual language modelling head. Used for implicit reward computation
        and loss. If no reference model is provided, the trainer will create a reference model with the same
        architecture as the model to be optimized.
    reward_funcs (`transformers.PreTrainedModel`):
        The reward model to score completions with, preferably an `AutoModelForSequenceClassification`.
    judge (`BasePairwiseJudge`):
        The judge to use for pairwise comparison of model completions.
    args (`NashMDConfig`):
        The NashMD config arguments to use for training.
    data_collator (`transformers.DataCollator`):
        The data collator to use for training. If None is specified, the default data collator
        (`DPODataCollatorWithPadding`) will be used which will pad the sequences to the maximum length of the
        sequences in the batch, given a dataset of paired sequences.
    train_dataset (`datasets.Dataset`):
        The dataset to use for training.
    eval_dataset (`datasets.Dataset`):
        The dataset to use for evaluation.
    processing_class ([`~transformers.PreTrainedTokenizerBase`], [`~transformers.BaseImageProcessor`], [`~transformers.FeatureExtractionMixin`] or [`~transformers.ProcessorMixin`], *optional*, defaults to `None`):
        Processing class used to process the data. If provided, will be used to automatically process the inputs
        for the model, and it will be saved along the model to make it easier to rerun an interrupted training or
        reuse the fine-tuned model.
    peft_config (`dict`):
        The peft config to use for training.
    compute_metrics (`Callable[[EvalPrediction], dict]`, *optional*):
        The function to use to compute the metrics. Must take a `EvalPrediction` and return a dictionary string to
        metric values.
    callbacks (`list[transformers.TrainerCallback]`):
        The callbacks to use for training.
    optimizers (`tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR]`):
        The optimizer and scheduler to use for training.
    preprocess_logits_for_metrics (`Callable[[torch.Tensor, torch.Tensor], torch.Tensor]`):
        The function to use to preprocess the logits before computing the metrics.

.. deprecated:: 0.22.0
    The following parameters are deprecated and will be removed in a future version:

    * `reward_model`: Use `reward_funcs` instead. For example, change `reward_model=model` to `reward_funcs=model`.

    c                 6  > Uc
  [        5       n[        USS5      n[        U5      [        La  Sn[        USS5      n[        U5      [        La  SnSn[        R
                  R                  SS5      S:H  nU(       d1  [        R
                  R                  SS5      S:X  a  [        S5        S	n[        R
                  R                  S
S5      n[        UR                  SS 5      =(       d    [        UR                  SS 5      nUc  UR                  5       R                  nSSKJn  U" U5      nU[        R                  :H  nU(       d  U(       a  U(       a  [        S5      eU(       d  U(       d  U(       a  [        S5      eU(       a"  SUl        SUl        S[        R
                  S'   OCU(       d<  U(       d5  US:X  a/  UUl        U(       + Ul        U(       a  SOS[        R
                  S'   [        USS 5      b-  [        USS5      S:X  a  SUl        [        USS 5      c  SUl        [        USS 5      nUb/  US:  a)  SSKJn  [-        U5      [-        S5      ::  a  [        S5        [        USS5      S:w  aL  [        USS5      nUS:X  a!  UR.                  U:  a  UR.                  Ul        [        US S 5      c
  Ub  UUl        [        US!S5      n[        U5      [        La  Sn[        US"S5      n[        U5      [        La  SnUR                   (       a  U(       a  SUl        S	Ul        UR"                  (       a  U(       a  S	Ul        SUl        U(       a  SUl        SUl        Oc[        R
                  R                  S
S5      S#:X  a  S	Ul        SUl        O0U(       d)  U(       d"  UR"                  Ul        UR                   Ul        Sn[9        5       R                  S$S 5      b  S	n[9        5       R                  S%S 5      b  S	nU(       a  S[        R
                  S&'   S'[9        5       ;  a  [;        US'5      (       d  OD[        US'S 5      n[        US'S 5      nUc'  Ub$  UR<                  n [;        US'5      (       a  U Ul        Ub!  [;        US(5      (       a  UR?                  5         S)[9        5       ;   a   [;        [@        S*5      (       a  S+[@        l!        S,[9        5       ;   aU  [;        U	S*5      (       a  S+U	l!        [;        U	S)5      (       a,  [;        U	R@                  S*5      (       a  S+U	R@                  l!        S,[9        5       ;   a  U	O[@        n!SS-K"J#n"  [I        UU"5      (       dx  [I        U[J        5      (       a(  S.URL                  ;  a  [O        U!SS/[        US0S 5      S19nO[I        U[N        5      (       a%  S.URL                  ;   a  [K        U![        US0S 5      S29nOJ[;        US35      (       a  SUl(        [;        US45      (       a  S5Ul)        [;        US65      (       a	  S7S	0Ul*        [I        UU"5      (       dx  [;        U!S85      (       dg  [;        U!S)5      (       aV  [I        U[J        5      (       a   [K        U!R@                  [        US0S 5      S29nO![O        U!R@                  SS/[        US0S 5      S19n/ n#SS9K+J,n$  U$" S:U#5        [        US;S 5      [Z        R\                  :X  a(  UR^                  S:  a  [        US<S5      S:w  a  SUl0        S=[9        5       ;   a!  [;        US(5      (       a  UR?                  5         [b        T'U ]  " SDUUUUUUUUU	U
UUUUS>.UD6  S=[9        5       ;   a!  [;        US?5      (       a  URg                  5         [;        U S@5      (       a-  U Rh                  Rk                  5         [;        U S@5      (       a  U ?4[        USAS 5      b  U Rl                  UR                  5       l6         [;        U SB5      (       aV  U Rn                  Rp                  n%Un&[;        U&S=5      (       a&  U%U&l9        U&Rt                  n&[;        U&S=5      (       a  M&  U%U&l9         [;        U SC5      (       a.  [w        [y        U Rz                  R|                  5      U 5      U l>        g )ENr   Fr   UNSLOTH_ENABLE_FULL_FINETUNING01UNSLOTH_FORCE_FLOAT32zKUnsloth: Switching to float32 training since model cannot work with float16TUNSLOTH_MIXED_PRECISIONr]   dtypetorch_dtyper   )
_get_dtypezuUnsloth: Model is in float16 precision but you want to use bfloat16 precision. Set fp16 to `True` and bf16 to `False`zuUnsloth: Model is in bfloat16 precision but you want to use float16 precision. Set fp16 to `False` and bf16 to `True`r   ACCELERATE_MIXED_PRECISIONro  r   r   r   r   r   rW   )__version__z4.45.2z**** Unsloth: Please use our fixed gradient_accumulation_steps by updating transformers, TRL and Unsloth!
`pip install --upgrade --no-cache-dir --force-reinstall --no-deps unsloth transformers trl unsloth_zoo`r      r   r   r   bfloat16rr  ru  UNSLOTH_RETURN_LOGITSr   r;   	tokenizerpadding_siderightrp  )UnslothVisionDataCollatorlabelsr   pad_to_multiple_of)mlmmlm_probabilityrX  )rX  r   dataset_text_fieldr   dataset_kwargsskip_prepare_datasetpad)PatchRLStatisticsnash_md_trainerparallel_mode_n_gpur:   )r:   rk  rl  r6  r?   rm  rn  ro  rp  rq  rr  rs  ru  rv  r<   neftune_hook_handler-  r  r  rS  )?r   getattrtypeboolr)   r/  getrT  r  get_input_embeddingsrK  unsloth_zoo.utilsrM  r,   float16	TypeErrorr   r   r   r   transformersrO  r2   r   r   r   r   r   localsr=   r   r;   rS  rT  unsloth_zoo.vision_utilsrV  r  r4   column_names+TransformersDataCollatorForLanguageModelingr   r[  r\  unsloth_zoo.logging_utilsr_  r6   NOT_DISTRIBUTEDr  rb  rY  rZ  r<   rc  remover-  r  scaleraccelerator_scalerr:   r7   rI   r[  r  )(r>   r:   rk  rl  r6  r?   rm  rn  ro  rp  rq  rr  rs  ru  rv  r@   use_bf16use_fp16force_float32full_finetuningmixed_precision_dtyperK  rM  rj  ga_stepstransformers_versioneval_bszr   r   _output_logitsmodel_max_seq_lengthargs_max_seq_lengthr    _UnslothNashMDTrainer__tokenizerrV  other_metricsr_  rt  current_modelr[  s(                                          rC   rZ  UnslothNashMDTrainer.__init__'  s   $ < 3 54/>%%x4/>%%x**..)I3OSVVBJJNN3JC$PTW$W_` M "

/H) Tgt4bm]a8b=%"<"<">"D"D%05!5==('hy  JA  @B  :Bg(9  NE  DF  >FDIDI7;BJJ3481F)1SDI#DIAHvfBJJ344.:wt_^b?cgk?k!(Dt\408C$/4!>EHqLH+,0AA @ A4$/47t%A1EH1}!A!AH!Lpt  qQ  qQdNmt6=E(J^  @H`d`| '7?t+e^ '7?t+e^99u)<\`dFY99t)<[`TEX"'D"'DZZ^^5yAZO"&D"'D"&))D"&))D8<<)40<tn8<<7>J]aN25BJJ./68+GDBR4S4S#*52BD#I #*42BD#I"*/C/O!&!5!54!122.D4G!?!? &("wy.'I'Idk9Ka)'88Za:J:W'55'BRB\B\^l:m:m  Zao  pJ  pJ  pW*<*H&iF-)BCC-)?@@XUbUoUoEo K&))07KT)R	! M+VWW\dhu  iC  iC  ]C 6)07KT)R!
 t455TYt7Qt122bD4Kt-..G]_cFd0C-)BCC;..7;3T3Tm-CDD$:#---4T;OQU-V%M
 %P#--#*--4T;OQU-V	%M ?+]; 4$/<3O3OOTXT^T^abTbtXq)Q.fh75.#A#A  	2!'))'/%-!,I'	2 +1	2 fh75/#B#B!4.//$$++-t2339Q4.5A?C?W?WE&&(<4''%%,,F!M-11390 - 3 3 -11 06M,4!!#$=dnn>R>R$SUYZDJrF   )r  )NNNNNNNNNNNNNN)r]  r^  r_  r`  ra  rZ  rd  re  rf  s   @rC   rD  rD    sA    .b (,k krF   rD  )Mra  r,   r   torch.nnr(   r   r   typingr   r   r   r   r	   r
   r   r   trl.trainer.nash_md_trainerr   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r    r!   r"   r#   r$   r%   r&   r'   r)   r*   r+   r-   r.   dataclassesr0   r1   packaging.versionr2   numpynp
contextlibr3   rl  r4   r5   rp  transformers.training_argsr6   rG   typesr7   rI   torch_compile_optionscompilerm   rc  rx   r   r   r   rh  rD  rS  rF   rC   <module>r     s  0    $ I I I {	  {	  {	  {	  {	  {	  {	  {	  {	  {	 
  ( %   " $  3      4;PR S"||  \\	&,, %  	
 \\6ell C ELL  p-, p- p-b tI, tIj[0 [z rF   