o ia@sdZddlZddlZddlZddlZddlZddlZddlmZm Z m Z ddl Z zddl Z dZWn ey;dZYnwdZdZdZd Zd d Zd d ZddZd4ddZddZGdddeZGdddZGdddeedZddZGdddeedZGd d!d!eZd"d#Z Gd$d%d%eZ!Gd&d'd'eZ"Gd(d)d)eZ#Gd*d+d+eZ$d,d-Z%Gd.d/d/ej&Z'Gd0d1d1e'Z(Gd2d3d3e'Z)dS)5zSpecifications declare the expected variables layout of CTranslate2 models that do not load a computation graph. The model converter should make sure that each required variable of the specification is set. N)DictListOptionalTF __optional)int8 int8_float32 int8_float16 int8_bfloat16int16float16bfloat16float32)rotary_scaling_long_factorrotary_scaling_short_factorcCs|s|Sd||fS)Nz%s/%s)scopenamerrY/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/ctranslate2/specs/model_spec.py _join_scope)s rcCs |dS)N/)splitrrrr _split_scope/ rcCs,t|}|dd|d}}d||fS)Nr)rjoin)rkeysattrrrr _parent_scope3src Cst|jD]@\}}|drqt|tr.t|D]\}}t||t|d||fdqqt|tr>t||t||dq||t|||qdS)z Recursively visits a layer spec._%s_%drN) list__dict__items startswith isinstance enumerate visit_specr LayerSpec)specfnrrvalueielemrrrr)9s   r)c Cs`|s|St|}|D]#}zt||}Wq ty-|dd\}}t||t|}Yq w|S)Nr!)rgetattrAttributeErrorrsplitint)r+indexrkeyrrrr index_specGs r7ceZdZfddZZS) FrozenMetacstj|i|}d|_|S)NT)super__call___frozen)selfargskwargsinstance __class__rrr;UszFrozenMeta.__call__)__name__ __module__ __qualname__r; __classcell__rrrArr9Tr9cr8) FrozenAttrcs2t|drt||std|t||dS)Nr<zAttribute %s does not exist)hasattrr2r: __setattr__r=r6r-rArrrJ\s zFrozenAttr.__setattr__)rCrDrErJrFrrrArrH[rGrHc @speZdZdZdddZ  dded edeeej ffd d Z d d Z ddZ dde eddfddZddZdS)r*zPA layer specification declares the weights that should be set by the converters.returnNcs4gfdd}||rtdddS)zVerify that the required weights are set. Raises: ValueError: If a required weight is not set in the specification. cs|dur |dSt|tjr|jtjkr|tj}n0t|tr,td |}n"t|t r:td |}nt|t rN|t krNtj |dtjd}t|tjsZt|tjr_t|}n trkt|tjrkt|}t|d}t|||dS)Nrrutf-8)dtyper)appendr'npndarrayrNfloat64astyperfloattypeboolstrOPTIONAL frombufferencodergeneric NumpyVariabletorch_is_availabletorchTensorPyTorchVariablersetattr)r+rr- attr_nameunset_attributesrr_checkms(         z"LayerSpec.validate.._checkz/Some required model attributes are not set: %s N)_visit ValueErrorr)r=rerrcrvalidatees  zLayerSpec.validater Fprefixorderedcs<ifdd}|||rttdddSS)a!Recursively returns the weights from this layer and its children. Arguments: prefix: Prefix to prepend to all variable names. ordered: If set, an ordered list is returned instead. Returns: Dictionary mapping variables name to value. cs(t|tr |tkr dS|t|<dSN)r'rWrXr)r+rr-rjvarrr _register_varsz*LayerSpec.variables.._register_varcSs|dSNrr)xrrrsz%LayerSpec.variables..)r6)rgr#sortedr%)r=rjrkrorrmr variabless  zLayerSpec.variablesc Cs||jdd}t|D]1\}}|D]*\}}||krn!t|\}}|s:||r:|tvr:t||}t|||nqq dS)z4Find duplicate variables in spec and create aliases.TrkN)rtreversedr is_scalarequalSKIP_CREATING_ALIASr7ra) r=rtrr- other_name other_valuerrbr+rrr_alias_variabless     zLayerSpec._alias_variablescs@durtvrtddtffdd}||dS)z-Possibly quantizes the variable of the layer.Nz;%s is not a valid quantization type. Accepted types are: %s, c st|tr |r dSt|d}d}t|d|}|jdv}|rÈdkrb|d}t dt t |}||9}t |}t |ttjjttjj}|tj}t|}t|}ndvr|d}d}t|jdkr|j}||jd d}tj t |d d }d ||d k<d |}|t|d 9}t |}|tj}|r||}t|}t|}n)d vr|}n|rdvr|d}ndvr|d}n dvr|d}t||||durt|d||dSdS)Nrz%s_scale)rr r r ri)rrr r rr0)axisg_@)r r r)r r r )r r r )rr r)r'VariablerwrrIrNtonumpyrPramaxabsoluterintclipiinfor minmaxrSr\lenshapereshape expand_dimsrra) r+rr-r6scaleis_quantizableis_convertible old_shaper quantizationrr _quantizes`               z&LayerSpec._quantize.._quantize)ACCEPTED_MODEL_TYPESrhrrg)r=rrrrrrs  =zLayerSpec._quantizercCs|||dS)aRRecursively applies some optimizations to this layer: * Alias variables with the same shape and value. * Quantize weights. Arguments: quantization: Weight quantization scheme (possible values are: int8, int8_float32, int8_float16, int8_bfloat16, int16, float16, bfloat16, float32). N)r|r)r=rrrroptimizes zLayerSpec.optimizecCst||dS)z/Recursively visits this layer and its children.N)r))r=r,rrrrgszLayerSpec._visitrLN)r Frl)rCrDrE__doc__rirWrVrrPrQrtr|rrrrgrrrrr*bs  ,  G r*) metaclassc Cs6d}z||WStytd|d|fw)N)rrr int32r r z)%s is not in list of supported dtypes: %sr})r5rhr) object_dtypedtypesrrr_dtype_to_type_ids   rc@s0eZdZdZddZddZddZdd Zd S) ModelConfigz$Base class for model configurations.cKs"|D] \}}t|||qdS)z7Initializes the configuration with a set of parameters.N)r%ra)r=r?r6r-rrr__init__'szModelConfig.__init__cCsdd|jDS)z*Returns the configuration as a dictionary.cSs i|] \}}|ds||qS)r!)r&).0r6r-rrr .s z'ModelConfig.to_dict..)r$r%r=rrrto_dict,szModelConfig.to_dictcCs||j|<dSrl)r$rKrrr add_attribute4zModelConfig.add_attributecCsTt|ddd}tj||ddd|dWddS1s#wYdS) z'Saves the configuration as a JSON file.wrMencodingT)indent sort_keysrfN)openjsondumprwrite)r=path config_filerrr save_as_json7s "zModelConfig.save_as_jsonN)rCrDrErrrrrrrrrr$s  rc@szeZdZdZddZeddZeddZedd Zd d Z dd e de e dd fddZ de dd fddZ ddZd S) ModelSpecz"The top level layer specification.cCs||_i|_dS)z$Initializes the model specification.N)get_default_config_config_filesrrrrrF  zModelSpec.__init__cCt)z$The name of the model specification.NotImplementedErrorrrrrrKzModelSpec.namecCdS)zThe model specification revision. This value is incremented each time the weights layout of the model is changed (e.g. a weight is renamed). r0rrrrrrevisionPszModelSpec.revisioncC|jS)zThe model configuration.)rrrrrconfigYrzModelSpec.configcCr)z5Returns the default configuration used by this model.Nrrrrrr^szModelSpec.get_default_configNrfilenamerLcCsPtj|s td||durtj|}||jvr!td|||j|<dS)z4Registers a file to be saved in the model directory.zFile %s does not existNz*A file with name %s was already registered)osrisfilerhbasenamer)r=rrrrr register_filebs     zModelSpec.register_file output_dircCs~|tj|d|jdur|jtj|d|jD]\}}tj||}tj|r6t d|t ||qdS)zwSaves this model on disk. Arguments: output_dir: Output directory where the model is saved. z model.binNz config.jsonz-File %s already exists in the model directory) _serializerrrrrrr%exists RuntimeErrorshutilcopy)r=rrr destinationrrrsavels  zModelSpec.savec s|g}g}|jddD]}t|dtr||q ||q t|dfdd}tdt||j td|j tdt ||D]A\}}||tdt |j |j D] }td|qgtdt |jtd||qPtdt ||D] \} } || || qWd d S1swYd S) zSerializes the model variables.Trur0wbcs@tdt|d|dtdddS)NHr0rMBr)rstructpackrrZ)stringmodelrr _write_stringsz+ModelSpec._serialize.._write_stringIrN)rtr'rWrOrrrrCURRENT_BINARY_VERSIONrrrrrrN num_bytesto_bytes) r=rrtaliasesvariablerrr-dimalias variable_namerrrr~s4         "zModelSpec._serializerl)rCrDrErrpropertyrrrrrWrrrrrrrrrCs    rccs\|D]&\}}t|dkr||dfVqt|D]\}}d||df|fVqqdS)Nr0rr")r%rr() vocabulariesr vocabularyr.vocabrrr_flatten_vocabulariess rc sLeZdZdZ      ddededed eed ed ef fd d ZZS)SequenceToSequenceModelConfigz.Configuration for sequence-to-sequence models.F unk_token bos_token eos_tokendecoder_start_tokenadd_source_bosadd_source_eosc s$tjd||||||d|dS)aInitializes the configuration for sequence-to-sequence models. Args: unk_token: The unknown token. bos_token: The start of sentence token. eos_token: The end of sentence token. decoder_start_token: The decoder start token. If ``None``, the token should be passed by the user in the target prefix. add_source_bos: If ``True``, ``bos_token`` will be automatically added to the source input. add_source_eos: If ``True``, ``eos_token`` will be automatically added to the source input. **kwargs: Additional configuration. )rrrrrrNrr:r)r=rrrrrrr?rArrrs z&SequenceToSequenceModelConfig.__init__)rrrrFF) rCrDrErrWrrVrrFrrrArrs*rcseZdZdZfddZddZejddZejdd Z d e e d d fd dZ d e e d d fddZ de d d fddZdfdd Zde d d ffdd ZZS)SequenceToSequenceModelSpecz3Base specification for sequence to sequence models.cstggd|_dS)z7Initializes a sequence to sequence model specification.sourcetargetN)r:r _vocabulariesrrArrrs  z$SequenceToSequenceModelSpec.__init__cCtSrl)rrrrrrz.SequenceToSequenceModelSpec.get_default_configcCr)z9Returns the source vocabulary size expected by the model.rrrrrget_source_vocabulary_sizerz6SequenceToSequenceModelSpec.get_source_vocabulary_sizecCr)z9Returns the target vocabulary size expected by the model.rrrrrget_target_vocabulary_sizerz6SequenceToSequenceModelSpec.get_target_vocabulary_sizetokensrLNcC|jd|dS)znRegisters a source vocabulary of tokens. Arguments: tokens: List of source tokens. rNrrOr=rrrrregister_source_vocabularyz6SequenceToSequenceModelSpec.register_source_vocabularycCr)znRegisters a target vocabulary of tokens. Arguments: tokens: List of target tokens. rNrrrrrregister_target_vocabularyrz6SequenceToSequenceModelSpec.register_target_vocabularyrcCs||ddS)zvRegisters a vocabulary mapping file. Arguments: path: Path to the vocabulary mapping file. zvmap.txtN)r)r=rrrrregister_vocabulary_mappingsz7SequenceToSequenceModelSpec.register_vocabulary_mappingcst||d}|D]H\}}t|ts|g}|j|}t|t|kr8t d|t|t|ft t ||D]\}\}}t||krYt d| |t||fq?qdS)NrzCIncorrect number of %s vocabularies: %d registered, but expected %dzK%s vocabulary %d has size %d but the model expected a vocabulary of size %d) r:rirrr%r'r#rrrhr(zip capitalize)r=vocabulary_sizesrsizesrr.r expected_sizerArrris0    z$SequenceToSequenceModelSpec.validatercsntt|j}t|tfddDrddi}|D] \}}t|d||q"t |dS)Nc3s|] }|dkVqdS)rNr)rrall_vocabulariesrr sz3SequenceToSequenceModelSpec.save..sharedrz %s_vocabulary) dictrrr#valuesallr%_save_vocabularyr:r)r=rrrrrArrrs  z SequenceToSequenceModelSpec.saver)rCrDrErrrabcabstractmethodrrrrWrrrrirrFrrrArrs   rcs6eZdZdZ   d dededeffdd ZZS) LanguageModelConfigz"Configuration for language models.rrrrrrc stjd|||d|dS)aInitializes the configuration for language models. Args: unk_token: The unknown token. bos_token: The start of sentence token. eos_token: The end of sentence token. **kwargs: Additional configuration. )rrrNrr)r=rrrr?rArrr+s zLanguageModelConfig.__init__)rrr)rCrDrErrWrrFrrrArr(srcspeZdZdZfddZddZejddZde e d d fd d Z dfd d Z de d d ffdd Z ZS)LanguageModelSpecz'Base specification for language models.cstg|_dS)z+Initializes a language model specification.N)r:r _vocabularyrrArrrErzLanguageModelSpec.__init__cCrrl)rrrrrrJrz$LanguageModelSpec.get_default_configcCr)z2Returns the vocabulary size expected by the model.rrrrrget_vocabulary_sizeMrz%LanguageModelSpec.get_vocabulary_sizerrLNcCst||_dS)zbRegisters the vocabulary of tokens. Arguments: tokens: List of tokens. N)r#rrrrrregister_vocabularyRsz%LanguageModelSpec.register_vocabularycs:t|}t|j|krtdt|j|fdS)NzEVocabulary has size %d but the model expected a vocabulary of size %d)r:rirrrrh)r=expected_vocabulary_sizerArrriZs  zLanguageModelSpec.validatercst|d|jt|dS)Nr)r rr:r)r=rrArrrdszLanguageModelSpec.saver)rCrDrErrrr rrrrWrrirrFrrrArrBs   rcCsVtj|d|}t|ddd}tj||ddWddS1s$wYdS)Nz%s.jsonrrMrr)r)rrrrrr)rrrvocabulary_pathvocabulary_filerrrr ls"r c@seZdZdZeejdeefddZ de fddZ eejde fddZ d e ddfd d Zejdejfd d Zde fddZejdefddZejdefddZejd e ddfddZejde fddZdS)rz(Abstract base class for model variables.rLcCrrlrrrrrrvrzVariable.shapecCst|jdkSrp)rrrrrrrw{rzVariable.is_scalarcCrrlrrrrrrN~rzVariable.dtyperNcCs||jkr|S||Srl)rN_tor=rNrrrrs  z Variable.tocCrrlrrrrrrzVariable.numpycCst|t|uo ||Srl)rU_equalr=otherrrrrxszVariable.equalcCrrlrrrrrrrzVariable.num_bytescCrrlrrrrrrrzVariable.to_bytescCrrlrrrrrrrz Variable._tocCrrlrrrrrrrzVariable._equalN)rCrDrErrr rrr4rrVrwrWrNrrPrQrrxrbytesrrrrrrrrss*rc@seZdZdZddZedeefddZede fddZ de j fd d Z defd d Zdefd dZde defddZdefddZdS)r\z Model variable as a Numpy array.cCs ||_dSrlarray)r=rrrrrrzNumpyVariable.__init__rLcC|jjSrl)rrrrrrrszNumpyVariable.shapecCs |jjjSrl)rrNrrrrrrNs zNumpyVariable.dtypecCrrlrrrrrrrzNumpyVariable.numpycCr rl)rnbytesrrrrrszNumpyVariable.num_bytescCs |jSrl)rtobytesrrrrrrzNumpyVariable.to_bytesrNcCsB|dkrts tdt|j|St|}|j||_|S)Nr z5Converting to bfloat16 requires torch to be installed) r]rr` from_numpyrrrPrNrSrrrrrs zNumpyVariable._tocCsL|j}|j}||up%|j|jko%|j|jko%|jd|jdko%t||Srp)rrNrflatrP array_equalr=rabrrrrs   zNumpyVariable._equalN)rCrDrErrrrr4rrWrNrPrQrrrrrrrVrrrrrr\s r\c@seZdZdZddZeddZedee fddZ ede fd d Z de jfd d Zde fd dZdefddZde defddZdefddZdS)r`z#Model variable as a PyTorch tensor.cCs"t|tjjr |j}||_dSrl)r'r^nn Parameterdata contiguoustensor)r=r-rrrrszPyTorchVariable.__init__cCst|}||Srl)r^r#)clsrr-rrrr#s zPyTorchVariable.from_numpyrLcCs t|jjSrl)r#r-rrrrrrs zPyTorchVariable.shapecCst|jjddS)Nztorch.r )rWr-rNreplacerrrrrNszPyTorchVariable.dtypecCs|jSrl)r-detachrrrrrrrzPyTorchVariable.numpycCs|j|jSrl)r-numel element_sizerrrrrszPyTorchVariable.num_bytescCsfd}|}d}d}|dkr1||kr|n|}t|j||}||7}||7}||8}|dks|S)Nir)rctypes string_atr-data_ptr)r=max_sizeroutputoffset chunk_sizechunkrrrrszPyTorchVariable.to_bytesrNcCstt|}|j||_|Srl)r1r^r-rrrrrrs zPyTorchVariable._tocCs,|j}|j}||up|j|jkot||Srl)r-rNr^rxr&rrrrs zPyTorchVariable._equalN)rCrDrErr classmethodr#rrr4rrWrNrPrQrrrrrrrVrrrrrr`s  r`)r )*rr r4rrrrtypingrrrrrPr^r] ImportErrorrXrrryrrrr)r7rUr9rHr*rrrrrrrrr ABCrr\r`rrrrsP    7 ^ &X*-.