+ jHj@Rt^RIHt^RIt^RIt^RIt^RIHtHt^RI H t ]P!] 4t Rt]P!R4t]P!R4t]P!R4t]P!R 4tR ]!R 04R ]!R 04/t]!4P,!]P/4!tR.tR Rlt/^R/b^R0b^R1b^R2b^R3b^R4b^R5b^ R6b^ R7b^ R8b^ R9b^ R:b^R;b^Rb^R?b/^R@b^R1b^R8b^RAb^RBb^RCb^R0b^R/b^RDb^RDb^REb^R0b^'RFb^(RGb^)RHb^*RIbCt^^ utt^R^R^R^R^R^R^R^R^ R^ R^ R/ tRRRRRRRRR R!R"R#R$R%/t]!R&R'44t R(R)lt!R*R+lt"R,R-lt#R#)JzGGUF metadata + tensor-table reader (stdlib only). Reads the header only (metadata + tensor infos); never touches tensor data, so it is fast enough to run at picker time on multi-GB files. ) annotationsN) dataclassfield)PathsGGUFz-(\d{5})-of-(\d{5})\.gguf$z-\d{5}-of-\d{5}$zblk\.(\d+)\.ffn_.*\.weightzblk\.(\d+)\.ssm_conv1d\.weightqwen4expzper_layer_token_embd.weightgemma4c V^8dQhRRRR/#)stemstrreturn)formats"B/home/ubuntu/.hermes/hermes-agent/hermes_cli/local_runtime/gguf.py __annotate__r's))S)S)c.\PRV4#)z(GGUFHeader.lazy_bytes..fs2H.Oldfn)/2F)F6.Os ** *)_LAZY_READ_TENSORSr2r4 frozensetsumr,items)r3r?s&@r lazy_bytesGGUFHeader.lazy_bytesbsK$''(9(99;GHd.G.G.M.M.OHH HrcV^8dQhRR/#)r suffixr r )rs"rrr0isBBBrc VVPPVP RV 24#).)r'r2r4r3rIs&&r _arch_keyGGUFHeader._arch_keyis)}}  D$5$5#6ax!@AArcV^8dQhRR/#r7r )rs"rrr0ms 6 6s 6rc VaRV3RllpV!R4V!R4r2SP'dV'd V'g^#V^V!R4,V,,p^W#,\^V!R4^, 4V,,,p\SP4V,#)zOne copy of one sequence's recurrent state: an f32 conv window plus the SSM state, per block that keeps state. llama.cpp holds one copy per slot per speculative draft position. 0 when the file declares no SSM shape.c V^8dQhRRRR/#)r keyr r r#r )rs"rr6GGUFHeader.recurrent_state_bytes..__annotate__qs : :S :S :rcR<\SPRV 24;'g^4#)zssm.r#rM)rRr3s&rssm-GGUFHeader.recurrent_state_bytes..ssmqs%t~~SEl388q9 9r inner_size state_size group_count conv_kernel)r-maxlen)r3rVinnerstateconv_dim per_blocksf rrecurrent_state_bytes GGUFHeader.recurrent_state_bytesls  : :<(#l*;u%%%%E1s=11E99QM0BQ0F)G()RRS 4(()I55rc V^8dQhRRRR/#)r rIr docr )rs"rrr0{sPP#PCPrc &a\V3RlVR7#)cL<\VPS4;'g^4#)rUrLs&r&GGUFHeader._arch_int..|sS)?)D)D1%Er)re)property)rIresf&r _arch_intGGUFHeader._arch_int{sE3OOr block_countcontext_lengthembedding_lengthzattention.sliding_window expert_countfull_attention_intervalz|GDN-hybrid discriminator (qwen35 family): every Nth layer is full attention, the rest are linear/recurrent. 0 = not present.zattention.key_length_swazPer-token key size for sliding-window layers when it differs from the global attention.key_length (e.g. gemma3/gemma4). 0 = not present.zattention.value_length_swazwPer-token value size for sliding-window layers when it differs from the global attention.value_length. 0 = not present.cV^8dQhRR/#)r r zlist[int] | Noner )rs"rrr0sDD(8Drc VPR4p\V\4'dVUu.uFp\V4NK up#R#uupi)uFPer-layer SWA pattern declared by the file itself: a truthy entry marks a sliding-window layer, falsy marks global. None when the file doesn't declare the per-layer array form (older GGUFs, architectures without per-layer SWA metadata, or files that use the scalar period form instead — see `sliding_window_pattern_period`). attention.sliding_window_patternN)rM isinstancelistr#)r3vxs& rsliding_window_pattern!GGUFHeader.sliding_window_patterns@ NN= >&0D&9&9"1A"CtC"sAcV^8dQhRR/#r7r )rs"rrr0sOOsOrc VPR4p\V\4'd"\V\4'g \V4#^#)uBScalar SWA period declared by the file: llama.cpp also permits `attention.sliding_window_pattern` as a single integer N (every Nth layer is full attention, e.g. Gemma-family writers) rather than a per-layer array. 0 when absent or when the file uses the array form instead — callers should fall back to a coarser signal.ru)rMrvr#boolr3rxs& rsliding_window_pattern_period(GGUFHeader.sliding_window_pattern_periods: NN= >#As++Jq$4G4Gs1vNQNrcV^8dQhRR/#r7r )rs"rrr0s:::rc VPR4pV'd \V4#VPPR4p\ V\ 4'd \ V4#^#)zrVocabulary size (prices the GPU logits buffers): vocab_size metadata when present, else the tokenizer list length. vocab_sizeztokenizer.ggml.tokens)rMr#r'r2rvrwr])r3rxtokss& rn_vocabGGUFHeader.n_vocabsN NN< ( q6M}}  !89&tT22s4y99rcV^8dQhRR/#)r r r&r )rs"rrr0s4rc /pVPP4Fwr#VPR4'gK\P VP R^4R,4pVfKP\ V\\34'gKn\\V4^4pV\V48Xd\\V44M \V4W&K V#)uEUpstream's recommended sampling as preset INI keys, when the file carries it. Publishers bake ``general.sampling.*`` keys into the GGUF (llama-server reads them as that model's defaults), so the file is the source of truth — it ships with the download and updates with every re-upload, no catalog needed. Empty if absent. zgeneral.sampling.rK) r'rE startswith_SAMPLING_INI_KEYr2rsplitrvr#floatroundr )r3outrRvaluer=nums& rsampling_defaultsGGUFHeader.sampling_defaultss----/JC>>"566$((C);B)?@DJusEl$C$CE%L!,-0CH_CCM#c( 0 rcV^8dQhRR/#r7r )rs"rrr0src VPR4p\V\4'd\\ V44#\T;'g^4#)zattention.head_count)rMrvrwr#r\rs& rn_headGGUFHeader.n_heads< NN1 2 a  s1v; 166{rcV^8dQhRR/#)r r z list[int]r )rs"rrr0s'' 'rc VPR4p\V\4'dVUu.uFp\V4NK up#\T;'g^4pVPpV^8d=\ VP 4Uu.uFpV^,V,^8XdTM^NK up#V.VP ,#uupiuupi)uqPer-layer KV head counts; 0 marks a recurrent/linear layer (n_head_kv == 0). Three GGUF shapes: a per-layer array (nemotron_h_moe) is used as-is; a scalar plus ``full_attention_interval`` (qwen35) applies to every N-th layer (1-indexed) and is zero elsewhere — pricing all layers as attention was a 4x overestimate; a plain scalar (dense) broadcasts to every layer. zattention.head_count_kv)rMrvrwr#rrrangen_layer)r3rxryscalarintervalis& rhead_counts_kvGGUFHeader.head_counts_kvs NN4 5 a  $%&AqCFA& &QVV!// a<"4<<020 !1u0A5F1<02 2x$,,&& '2s B;"CcV^8dQhRR/#r7r )rs"rrr0s@@C@rc VPR4pV'd \V4#VP'dVPVP,#^#)zattention.key_length)rMr#rn_embdrs& r head_dim_kGGUFHeader.head_dim_ks= NN1 2 q6M-1[[[t{{dkk)?a?rcV^8dQhRR/#r7r )rs"rrr0sCrc bVPR4pV'd \V4#VP#)zattention.value_length)rMr#rrs& r head_dim_vGGUFHeader.head_dim_vs' NN3 4 q6Mrr N)r)&__name__ __module__ __qualname____firstlineno____annotations__rr&r'r(r)r*r+r,rCr-rkr4rFrMrbrlr n_ctx_trainrsliding_windowrqrrkey_length_swavalue_length_swarzrrrrrrr__static_attributes__r rrr!r!Js I L40Hd0IsL#c',D&AO^A,1+F.F(1{n2BBHH B 6 6P &G,-K ) *F9:N^,L'! 12" FGN!$ 34  DDOO::" '$@@ rr!c V^8dQhRRRR/#)r r"rr zlist[Path] | Noner )rs"rrrs - -d -0 -rcaaa\PSP4pVfR#SPRVP4\ VP ^44uooVVV3Rl\ ^S^,44Uu.uFpVP4'gKVNK pp\V4^8dV#R#uupi)aEvery on-disk part of the split ``path`` belongs to, first part first; None when ``path`` is not a split member or no other part is present. A split is priced as the set, never as one file: publishers lay shards out so the first part can hold little more than metadata while the bulk sits in the later ones, so reading one part prices the model at whatever fraction of its weights that part happens to hold.Nc 3\<"TF!pSPS RVR RSR R24xK# R#5i)-05dz-of-z.ggufN) with_name)r<rr"r totals& rr@split_parts..s;7"5Q4&!CU3Ku(MNN"5s),) SPLIT_PART_REsearchr=startr#groupris_filer])r"mppartsr rsf @@r split_partsrs TYY'Ay))Kaggi(#aggaj/KD%7"'519"57 G71:;))+Q7E GJN5,, Gs B;B;c V^8dQhRRRR/#)r r"rr r!r )rs"rrrs"DDDDTDDjDDrc aaaaRV3RlloRV3RlloRVVV3Rllo\SR4;_uu_4pVP^4\8wd\RS 24hS!VR 4wr#p/p\ V4F pS!V4pS!VR 4wpS!W4WW&K" ^p ^p /p /p \ 4p \ V4EFWpS!V4pS!VR 4wpS!VR V R 24pS!VR 4wpVP^4\ PV4pVf\RV RS 24hVwpp^pVF pVV,pK VV,V,pV V, p VR8XdTp K\PV4;p'd9\VP^44pV PV^4V,V V&KV\9dVW&EK\PV4;p'gEK.V P\VP^444EKZ R R R 4\\!S4XXXX X X X \#X 4R7 # +'giL6;i)z=One file's own header: metadata and that file's tensor table.cV^8dQhRR/#)r fmtr r )rs"rr _read_part..__annotate__s((S(rc<\P!V4pVPV4p\V4V8wd\ RS 24h\P !W4#)ztruncated GGUF header: )structcalcsizereadr] ValueErrorunpack)frsizedatar"s&& rr_read_part..readsNs#vvd| t9 6tf=> >}}S''rcV^8dQhRR/#r/r )rs"rrrs;;s;rc^<S!VR4wpVPV4PRRR7#)rzutf-8replace)errors)rdecode)rnrs& rread_str_read_part..read_strs.At}vvay ::rcV^8dQhRR/#)r vtyper#r )rs"rrrs..S.rc<V\8Xd S!V4#V\8Xd-S!VR4wr#\V4Uu.uF pS!W4NK up#S!V\V,4^,#uupi)z.read_valuesd I A;  H Au~HE27(;(QJq((; ;A{5)*1--rblockrrrsf @@@r _read_partrsH((;;.. dD  Q 66!9 #078 8#'6? DtA1+CAt}HU&q0HM   *,/1%(Uy!AA;DQ IVQvha=)DAt}HU FF1I#''.D| #s+U+U:+U*+Urc|\V4p\V4pVf \V4#\V^,4pV.pVR,FpVP\V44K \V4^8XdV#/pVFApVPP4F wrVPV^4V ,Wh&K" KC \\V4VPVP \#RV44\#RV44\#RV44TVU U U u/uF(p V P$P4FwrWbK K* up p p \'4P(!RV4!R7 # \\ 3d.p\ PRTPT4Rp?EKkRp?ii;iuup p p i) uHeader for a MODEL, not for a file: a split GGUF is priced as the sum of its parts. Weights, the host-side embedding-table duplicate and the per-block FFN map are all summed across the shards on disk, because a split is a layout choice, not a smaller model — pricing part 1 alone underprices every model whose first shard is a metadata stub (the Hugging Face layout), which then makes the physics check and the residency cap admit giants the card cannot hold. Architecture metadata (block count, train context, the per-layer SWA pattern, vocab) is taken from the first part: gguf-split writes the model's metadata there only, and every later part carries just its ``split.*`` keys. A part that has gone missing or become unreadable is skipped rather than fatal: a half-arrived split prices at what is actually on disk. Refusing it is ``staged_in(require_complete=True)``'s job — an incomplete split is never servable, it is only underpriced here.N:NNzsplit part unreadable %s: %sc38"TFqPxK R#5ir:)r(r<hs& rr@#read_gguf_header..cs48akk8c38"TFqPxK R#5ir:)r)rs& rr@rds:Arc38"TFqPxK R#5ir:)r*rs& rr@resBA//rc38"TFqPxK R#5ir:)r-rs& rr@ris,RA-?-?rr)rrrappendrOSErrorloggerdebugr=r]r+rEr2r!r r$r'rDr,rCunion) r"rfirstreadablepartexcr+headerrr>rr=s & rread_gguf_headerr>s :D  E }$ uQx EwHb  I OOJt, -  8} &(O#3399;ME%4%8%8%BV%KO "< Y 4844:::BBB'4<RHq232H2H2N2N2P,$#l2P#HR"**,R,RS U UG$ I LL7C H H I"RsE6 .F76F4!F//F4l)r)r r) )r)r)r)"r)$r)T)nr)r)r)r)i$r)Br)Jr)br)2r)Rr)r)rr)r)8r)r)r@)r)rr,)$__doc__ __future__rloggingrer dataclassesrrpathlibr getLoggerrr rcompilerrrrrCrBrvaluesrr;rrrrrrr!rrrr rrr7s # (   8 $  89 **01jj67 JJ8 9  89: i678" ))+=+D+D+FG) v & X '(( 45x ABH x H  ) -/ =?     %'  57   y #%i 24X      $&z  46v  @B6    F    )+I  8:6       8tQaq$tQaq$b$D V]FGWgw%'7');=  YY Yx - DDN+Ur