+ "ݳj^RIt^RIt]!]P!RR44tRRltRRltRR ltR R R ]/R RlltR ]/RRllt R#)N!GPTQMODEL_NPU_CHOLESKY_BLOCK_SIZE64c(V^8dQhR\RR/#)messagereturnN)str)formats"b/home/luca/screen-checking/.venv/lib/python3.14/site-packages/gptqmodel/quantization/npu_linalg.py __annotate__r s))))c@\PPV4h)N)torch_C _LinAlgError)rs&r _raise_linalg_errorr s ((   ((r cHV^8dQhR\PR\RR/#)rmatrixop_namerN)rTensorr )r s"r r r s*ggELLg3g4gr cVP^8wg*VP^,VP^,8wd%\V R\VP4 R24hVP 4'g\ V RVP R24hVPPR8wd\V RVP R24hR#)rz' expects a 2D square matrix, got shape .z, expects a floating point matrix, got dtype npuz/ is only intended for NPU tensors, got device `z`.N) ndimshape ValueErrortupleis_floating_point TypeErrordtypedevicetype)rrs&&r _validate_square_matrixr#s {{a6<<?fll1o=G9$KERXR^R^L_K``abcc  # # % %7)#OPVP\P\~]^_`` }}U"G9$STZTaTaSbbdeff#r cXV^8dQhR\PR\P/#)rpanelr)rr)r s"r r r s"%,,r cVP^,p\P!V4p\V4EF(pV^8Xd WV3,pM7W#RV13,pWV3,\P!WU,4, p\ \P !V4V^8,4'g \R4\P!V4pWbW33&V^,V8gKV^8XdW^,R1V3,pMFW^,R1V3,W#^,R1RV13,PW#RV13,4, pWv, W#^,R1V3&EK+ V#)rNz@NPU Cholesky failed because the matrix is not positive-definite.) rr zeros_likerangesumboolisfinitersqrtmatmul)r%sizefactorcoldiag_sqrowdiag numerators& r _cholesky_panel_lowerr5s+ ;;q>D   e $FT{ !8HoGdsd#CHo #)(<ax!'(C-0 !'(C-06'(DSD.3I3P3PQW]a^a]aXaQb3cc $-$4F78S= !%( Mr upperF block_sizecpV^8dQhR\PR\R\R\P/#)rrr6r7r)rrr*int)r s"r r r 4s:-'-' LL-' -' -'  \\ -'r cP\VR4VP^,pV^8XdVP4#\^\ V44p\ P !V4p\^W24EFp\WR,V4pWV1WV13,P4pV'd0WEV1RV13,pWxPVP4, p\V4p WWV1WV13&Wc8gKWR1WV13,p V'd;WVR1RV13,PWEV1RV13,P4, p \ PPV V PP4RR7p V PWFR1WV13&EK" V'd VP#T#)a  Compute a Cholesky factor on NPU without aten::linalg_cholesky_ex. torch-npu 2.9 falls back to CPU for torch.linalg.cholesky on this path. This blocked implementation keeps the dense GPTQ Hessian factorization on NPU and uses only matmul, elementwise ops, and solve_triangular updates. npu_choleskyNFr6)r#rclonemaxr9rr'r(minr-Tr5linalgsolve_triangular contiguous) rr6r7r.lowerstartendr%previous panel_factortrailingsolveds &$$ r r;r;4s}FN3 <<?D qy||~QJ(J   V $Eq$+%$d+Sy%)+,224 3Y./HOOHJJ77E,U3 &2ei"# :dEIo.H#CD&5&L&9&@&@SyRXSXRXGXAYA[A[&\\\\22 %%'3F &,XXE$ / "+,.577&&r cdV^8dQhR\PR\R\P/#)rrr7r)rrr9)r s"r r r ds0** LL** \\ *r c\VR4\VRVR7p\P!VP^,VP VP R7p\PPW#RR7p\PPVRR7wrV\P!\P!V44p\P!V^8H\P!V4V4pWgP^4,#)a Return the upper Cholesky factor of inv(matrix) on NPU. This matches the GPTQ solver contract used by torch: torch.linalg.cholesky(torch.cholesky_inverse(torch.linalg.cholesky(matrix)), upper=True) while avoiding the torch-npu CPU fallbacks for cholesky and cholesky_inverse. npu_inverse_cholesky_factorF)r6r7)r r!r<reduced)mode)r#r;reyerr r!rArBqrsigndiagonalwhere ones_like unsqueeze)rr7rDidentity lower_inv_r6 diag_signs&$ r rMrMdsF$AB u DEyya V]]SH --eU-KI||yy9HA 5>>%01I INEOOI,F RI &&q) ))r ) osrr9getenv_DEFAULT_CHOLESKY_BLOCK_SIZErr#r5r;rMr r r_se  #299-PRV#WX)g6-'-'3 -'`*3**r