`i.W ddlZddlmZmZddlZddlmZddlm Z m Z m Z ddl m Z ddlmZddlmZe rddlZeeZd d d d d ddde de dide didZeedZGddeZGddZGddeZGddeZGddeZGdd eZGd!d"eZGd#d$eZ Gd%d&eZ!Gd'd(eZ"Gd)d*eZ#eeeeeeee e!e"e"e#d+ Z$d,Z% d4d.ee&d/ee'd0e&fd1Z(d5d3Z)dS)6N) NamedTupleOptional)tqdm)GGUF_CONFIG_MAPPINGGGUF_TOKENIZER_MAPPING_gguf_parse_value)is_torch_available)is_gguf_available) get_loggerversion tensor_countkv_count)r rr file_typequantization_version)rr)GGUFgeneral tokenizertokenizer_config)ignoreconfigrrrc8eZdZUejed<eed<eed<dS) GGUFTensorweightsnamemetadataN)__name__ __module__ __qualname__npndarray__annotations__strdict|/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/transformers/modeling_gguf_pytorch_utils.pyrr7s0 Z IIINNNNNr&rceZdZddZdZdS)TensorProcessorNc|pi|_dSNr)selfrs r'__init__zTensorProcessor.__init__>sl r&c $t||iSr+rr-rrkwargss r'processzTensorProcessor.processAs'4,,,r&r+)rrrr.r3r%r&r'r)r)=s7####-----r&r)c beZdZd fd ZdZ d dejdedeedejfdZ xZ S) LlamaTensorProcessorNcLt|dSNr,superr.r-r __class__s r'r.zLlamaTensorProcessor.__init__F$ '''''r&c 8d|vsd|vr|jd}|jd}d||fvrt||iSd|vr||||}nd|vr||||}t||iS)Nz.attn_k.z.attn_q.num_attention_headsnum_key_value_heads)rgetr_reverse_permute_weights)r-rrr2 num_heads num_kv_headss r'r3zLlamaTensorProcessor.processIs   t!3!3 (=>>I;??+@AAL <000!'4444T!!77IVVt##77LYY'4,,,r&rn_headrCreturnc|||kr|}|jd|zdz}|j||dg|jddR}|dd|jS)Nrr)shapereshapeswapaxes)r-rrDrCdimws r'rAz-LlamaTensorProcessor._reverse_permute_weightsVs{  #,(>(>!FmA&(A- GOFC ?W]122-> ? ? ?zz!Q'' 666r&r+) rrrr.r3r r!intrrA __classcell__r;s@r'r5r5Es(((((( - - -OS 7 7z 7+. 7>Fsm 7  7 7 7 7 7 7 7 7r&r5cXeZdZd fd ZdZdejdeeefdedefdZ xZ S) Qwen2MoeTensorProcessorNcLt|dSr7r8r:s r'r.z Qwen2MoeTensorProcessor.__init__dr<r&c d|vrU|d}|d}|r)|||||t|diSd|vrtj|d}t||iS)N_exptensor_key_mappingparsed_parametersffn_gate_inp_shexpraxis)r@_split_moe_expert_tensorrr expand_dims)r-rrr2rUrVs r'r3zQwen2MoeTensorProcessor.processgs T>>!',@!A!A  & +> ? ? ! 5--g7H$Pbccc!'4444 4 ' 'nW1555G'4,,,r&rrVrrUc||}|jdd}td|D]S}|dd|d}||}t jt j||d|<TdS)N num_experts<r mlp.experts..tensors)rr@rangereplacetorch from_numpyr copy) r-rrVrrU w_counteri temp_name exp_weights r'rZz0Qwen2MoeTensorProcessor._split_moe_expert_tensorts "$'KOOM266 q)$$ \ \A ^5HA5H5H5HIII J6;6FrwzGZGZ6[6[ i ( 3 3 \ \r&r+) rrrr.r3r r!r$r#rZrNrOs@r'rQrQcs(((((( - - - \z \6:39o \MP \fj \ \ \ \ \ \ \ \r&rQcbeZdZd fd ZdZdejdedefdZdejdedefdZ xZ S) BloomTensorProcessorNcLt|dSr7r8r:s r'r.zBloomTensorProcessor.__init__r<r&c d|vrM|jd}|jd}d|vr||||}n||||}t||iS)Nattn_qkvrD hidden_sizeweight)r_reverse_reshape_weights_reverse_reshape_biasr)r-rrr2rBn_embeds r'r3zBloomTensorProcessor.processsr    H-Ik-0G477GTT44WiQQ'4,,,r&rrDrtcDtj|dd\}}}||||z|}||||z|}||||z|}tj|||gd}||dz||zz|S)NrrXr)r array_splitrIstack)r-rrDrtqkv qkv_weightss r'rrz-BloomTensorProcessor._reverse_reshape_weightss.!!4441a IIfg/ 9 9 IIfg/ 9 9 IIfg/ 9 9h1ayq111 ""6A:F1B#CWMMMr&c$tj|d\}}}||||z}||||z}||||z}tj|||gd}|S)NrvrrX)r rwrIrxflatten)r-rrDrtq_biask_biasv_biasqkv_biass r'rsz*BloomTensorProcessor._reverse_reshape_biass"$!;!;6(9::6(9::6(9::8VVV41===EEGGr&r+) rrrr.r3r r!rMrrrsrNrOs@r'rlrls((((((--- N NC NRU N N N N RZ  s        r&rlc&eZdZdfd ZdZxZS)T5TensorProcessorNcLt|dSr7r8r:s r'r.zT5TensorProcessor.__init__r<r&c d}|dD]'}|rt|}n(t||d|iS)Nr`bid)splitisdigitrMr)r-rrr2rchunks r'r3zT5TensorProcessor.processs]ZZ__  E}} %jj '4%666r&r+rrrr.r3rNrOs@r'rrsL((((((7777777r&rc&eZdZdfd ZdZxZS)GPT2TensorProcessorNcLt|dSr7r8r:s r'r.zGPT2TensorProcessor.__init__r<r&c d|vs d|vsd|vsd|vr|j}|dkrId}|di}tjt j||d|<d}t ||iS) Nzattn_qkv.weightzffn_down.weightz ffn_up.weightzattn_output.weight output.weightzlm_head.weightrVra)Tr@rdrer rfr)r-rrr2rVs r'r3zGPT2TensorProcessor.processs  % % D(($&&#t++iG ? " "$D & +> C C 161A"''BRBR1S1S i ( .D'4,,,r&r+rrOs@r'rrsL((((((-------r&rc&eZdZdfd ZdZxZS)MambaTensorProcessorNcLt|dSr7r8r:s r'r.zMambaTensorProcessor.__init__r<r&c d|vrtj|d}d|vrtj| }t||iS)Nzssm_conv1d.weightrrXssm_a)r r[logrr1s r'r3zMambaTensorProcessor.processsQ $ & &nW1555G d??fgX&&G'4,,,r&r+rrOs@r'rrsL(((((( - - - - - - -r&rc&eZdZdfd ZdZxZS)NemotronTensorProcessorNcLt|dSr7r8r:s r'r.z NemotronTensorProcessor.__init__r<r&c 6d|vr|dz }t||iSNz norm.weightrr0r1s r'r3zNemotronTensorProcessor.process( D kG'4,,,r&r+rrOs@r'rrsL((((((-------r&rc&eZdZdfd ZdZxZS)Gemma2TensorProcessorNcLt|dSr7r8r:s r'r.zGemma2TensorProcessor.__init__r<r&c 6d|vr|dz }t||iSrr0r1s r'r3zGemma2TensorProcessor.processrr&r+rrOs@r'rrsL(((((( -------r&rc&eZdZdfd ZdZxZS)Lfm2TensorProcessorNcLt|dSr7r8r:s r'r.zLfm2TensorProcessor.__init__r<r&c Xd|vrtj|d}t||iS)Nzshortconv.conv.weightrrX)r r[rr1s r'r3zLfm2TensorProcessor.processs3 "d * *nW1555G'4,,,r&r+rrOs@r'rrsL((((((-------r&r) llamaqwen2moeqwen3moebloomt5 t5encodergpt2mambanemotrongemma2gemma3lfm2cZ||jvrgS|j|fdjDS)NcPg|]"}tj|j#Sr%r partstypes).0 _data_indexvalues r' zread_field..s- c c c ek+6 D D c c cr&)fieldsdata)readerfieldrs @r' read_fieldrs? FM!! M% E c c c cX]Xb c c ccr& model_type num_layers qual_namec trtr ddlm}m}n)t dtd| |jj n|}| |jj n|}|dkrd}n#|dkrd }n|d krd }n|d krd }n|dkrd}d}| D]\}}||kr|}n|td|d|||} i| } | D]} |dvrd| vrtjdd| } | d} } | ds| dr| dd\} } d| z} | | }||| z|| z<|x}rU|D]R\} }t)||||| d}fd| D}|SS)aY GGUF uses this naming convention for their tensors from HF checkpoint: `blk.N.BB.weight` and `blk.N.BB.bias` where N signifies the block number of a layer, and BB signifies the attention/mlp layer components. See "Standardized tensor names" in https://github.com/ggerganov/ggml/blob/master/docs/gguf.md for details. r)MODEL_ARCH_NAMESget_tensor_name_mapLoading a GGUF checkpoint in PyTorch, requires both PyTorch and GGUF>=0.10.0 to be installed. Please see https://pytorch.org/ and https://github.com/ggerganov/llama.cpp/tree/master/gguf-py for installation instructions.KPlease install torch and gguf>=0.10.0 to load a GGUF checkpoint in PyTorch.Ncoherez command-r qwen2_moer qwen3_moer gemma3_textrumt5rzUnknown gguf model_type: z in gguf-py. This might because you're using an outdated version of gguf-py package, you can install `gguf` package from source refer to https://github.com/ggerganov/llama.cpp/tree/master/gguf-py#development)rrr_zmlp.experts.\d+.rz.weightz.biasr`r)rc$i|] \}}|v || Sr%r%)rrzr{gguf_to_hf_name_maps r' z+get_gguf_hf_weights_map..es*XXX11DW;W;Wq!;W;W;Wr&)r r ggufrrloggererror ImportErrorrrnum_hidden_layersitemsNotImplementedError state_dictresubendswithrsplitget_namenamed_childrenget_gguf_hf_weights_mapupdate)hf_modelrrrrrarchkeyrname_maprhf_namersuffix gguf_namerchildsub_maprs @r'rrsi133i>>>>>>>>>  A   ghhh/9/A++zJ6@6H22jJX { " " { " " } $ $ v   D&,,.. U J  D E  |! U  U U U   #"444H$$&&JFF 1 1 1n6O6Of0.'JJGf   I & & "'*:*:7*C*C "">>#q11LD&6\F%%d++   2;g2EI.//"00222~0) 0 0KD%-eZZcWleiWlWlWlmmmGXXXX XXXG  & &w / / / / r&Fcd !"#trtr ddlm}m}n)t dtd||}|j}t| }dtD}t|dd} t|d} d} d | vrd | vrd } nYd | vsd | vrOd |dd<| r/d| d vrd} d | vr dg|dd<nd | vr dg|dd<d } n| } d| vrd} nd| vrd} d| vr]hd!d"t!fd|jD} t"fd|jD} | |dd<| |dd<| t vr| t vrt#d| d d!d"g}t%d#|jDp| |v|dd$<|jD]?\}#|| | }|d%}|d}d%|d&d}#fd'#jD}t1|d&kr|d}t3|t4r| |vr|| | }tD]N\}}||vrE|||vr;|||}|d(kr(| ||||<||vr||O||vr t d)|d*|A|dd+d,kr d-|dd+<|dd+d.krS|dd/}t;||dd/<d0|dd1<d2t=|D|dd3<d4|dvrE|d5}d6|vrt1|d6|dd4<nt d7|ri|d8<tA|}|!di}tD!| tF}||9}tI|jd:;D]}|j%}||j|j&}|'||||<} | j(}| j%}||vrK||}tSj*tWj,||d8|<t1|dkrt d=||S)>a Load a GGUF file and return a dictionary of parsed parameters containing tensors, the parsed tokenizer and config attributes. Args: gguf_checkpoint_path (`str`): The path the to GGUF file to load return_tensors (`bool`, defaults to `False`): Whether to read the tensors from the file and return them. Not doing so is faster and only loads the metadata in memory. r) GGUFReader dequantizerrci|]}|iSr%r%)rrzs r'rz(load_gguf_checkpoint..sEEE1BEEEr&zgeneral.architecturez general.nameNrmistralrrTr is_gated_actrUMT5EncoderModel architecturesT5EncoderModelrrrrstablelm> attn_k.bias attn_q.bias attn_v.biasffn_normc34K|]}D] }||jvVdSr+r)rtensor bias_nameattn_bias_names r' z'load_gguf_checkpoint..s9mmF^lmmQZyFK/mmmmmmmr&c3*K|] }|jvVdSr+r)rr ffn_norm_names r'rz'load_gguf_checkpoint..s*#^#^VMV[$@#^#^#^#^#^#^r& use_qkv_biasuse_parallel_residualzGGUF model with architecture z is not supported yet.falconrc3,K|]}d|jkVdS)rNr)rrs r'rz'load_gguf_checkpoint..s) H HvOv{ * H H H H H Hr&tie_word_embeddingsr`rcPg|]"}tj|j#Sr%r)rrrs r'rz(load_gguf_checkpoint..s-hhhk"5;{#;U[IIhhhr&z1Some keys were not parsed and added into account z | rrrrr?Fblock_auto_adjust_ff_dimc$g|] \}}|dk |S)rr%)rrhrCs r'rz(load_gguf_checkpoint..s.9 9 9 !!\L[\L\L\AL\L\L\r&full_attn_idxs vocab_sizertokenszCan't find a way to retrieve missing config vocab_size from tokenizer parameters. This will use default value from model config class and cause unexpected behavior.rar,z,Converting and de-quantizing GGUF tensors...)desc)rrrUrVz0Some keys of the GGUF file were not considered: )-r r rrrrrrrlistkeysGGUF_TO_TRANSFORMERS_MAPPINGrloweranyraGGUF_SUPPORTED_ARCHITECTURES ValueErrorallrrcrjoinrlen isinstancer#removeinfomax enumeratewarningrr@TENSOR_PROCESSORSr)rr tensor_typer3rrdrer rf)$gguf_checkpoint_pathreturn_tensors model_to_loadrrrr reader_keysrV architecture model_nameupdated_architecturerr exceptionsgguf_keyrprefix config_keyr parameterparameter_renamesrenamed_config_keygguf_num_key_value_headstokenizer_parametersrUrProcessorClass processorrrrresultrrrs$ @@@r'load_gguf_checkpointr2ksi133i/////////  A   ghhh Z, - -F ]Fv{{}}%%KEE(DEEEf&<==a@LFN33J,9 #:#:(    !(#N3CX?X(#$;<777z%J"%++%1GL%i01CD{**&&x000 { " " KK`H``Y^`` a a a"<0H<<4A(#L1"<0F::#4X#>?T#U =@AY=Z=Z(#$9:BG(#$>?9 9 %./G%H%H9 9 9 (#$45 ,X6660= + + +8;0^___ T TF;D jf.@AAG&&#5"3 'FnG;D---%d+D161A"''BRBR1S1S i ( . . ;! T{TTUUU r&)NNr)FN)*rtypingrrnumpyr tqdm.autor integrationsrrr utilsr utils.import_utilsr utils.loggingr rdrrrr rrrr)r5rQrlrrrrrrrrr#rMrr2r%r&r'r:s  '''''''' &%%%%%111111%%%%%%LLL H   !*"  "-F\]] "5kBC$&<=O&PQ   $t$@$J$O$O$Q$QRR --------77777?777<\\\\\o\\\<$$$$$?$$$N 7 7 7 7 7 7 7 7-----/---4 - - - - -? - - - -----o--- - - - - -O - - ------/---"'' ! "  !'##    ddd!% $ OO O O OOOOdiiiiiir&