+ "ݳjlaRt 0t^RIt^RIHt]!4tRs]^kRRltRR/RRlltRR R/R R llltR R lt RRlt Rt R#)N) setup_loggerc$V^8dQhR\/#)return)bool)formats"[/home/luca/screen-checking/.venv/lib/python3.14/site-packages/gptqmodel/quantization/gar.py __annotate__r s  $ c\f8\P!\P!RR.4RR7Rs\#\# \d Rs\ P R4\#i;i)Ngg?T)stableFz-Torch: missing argsort with `stable` support.)_HAS_STABLE_ARGSORTtorchargsorttensor TypeErrorlogwarnr r _supports_stable_argsortrsi" F MM%,,Sz24 @"&    F"'  HHD E  Fs0A"A0/A0 return_valuesFc V^8dQhR\PR\R\R\P\\P\P3,,/#)rdiag_H groupsizerr)rTensorintrtuple)rs"r r r &sO!! LL!! !  \\E%,, 455 !r cpVP4pW1,pV^8Xdl\P!^V\PVPR7pV'd0V\P!^WP VPR73#V#VRWA,P WA4pVP;'dV^8*;'dV^8pV'd\P!Wa^RRR7wrM\P!V^RR7wrV P\PR7p V'dW3#V #)zVectorized within-group permutations. Optionally returns the sorted diagonal statistics so callers can reuse them (e.g. to derive max-based group scores without another kernel launch). dtypedeviceNT)kdimlargestsorted)r# descendingr ) numelremptylongr!r viewis_cudatopksortto) rrrn num_groupsr)Huse_topkvaluesindicess &&$ r compute_local_permsr6&s  AJQ Ay 6==Q %++a,,v}}]] ] ''(--jDA~~DD)s"2DDy1}H**QDQUV**QA$?jjuzzj*G Nr precomputed_valuesc V^8dQhR\PR\R\R\PR,R\P/#)rrrmetricr7Nr)rrrstr)rs"r r r KsK)) LL)) )  t+ )  \\ )r cVP4pWA,pV^8Xd2\P!^\PVPR7#VR8XdVe VR ,pMVRWQ,P WQ4pVR8XdVP ^R7PpMfVR8XdVP^R7pMLVR8XdVP^R7pM2VR8XdVP^R7PpM\RV 24h\4'd\P!VR R R 7pV#\P!VR R 7pV#) u~ Returns a (num_groups,) LongTensor with the order of groups by a chosen metric. metric ∈ {"max", "mean", "sum", "median"}. rmaxNr#meansummedianzUnknown metric: T)r&r )r&):NNNr)r(rr)r*r!r+r<r4r>r?r@ ValueErrorrr) rrr9r7r0r1scoresr2 global_perms &&&$ r compute_global_permrDKs   AJQ{{1EJJv}}EE -9#D) +Z+ , 1 1* H U?UUqU\((F v VVV]F u_UUqU\F x XX!X_++F/x89 9 !!mmFtDI  mmFt< r cDV^8dQhR\R\P/#)rrr)rrr)rs"r r r wsCELLr c p\V\4'd[V^,Pp\P!VUu.uF#qDP V\P R7NK% up^R7pM VP \P R7pVPwrgWr8XgQRV RV 24h\P!WePR7V,PV^4pWX,VP VP\P R7,p V PR4#uupi)z Vectorized: - Stack local_perms once (if it's a list) -> (G, S) - Add per-group base offsets - Reorder groups by global_perm - Flatten Returns a permutation p with the SAME semantics as your original: p[new_pos] = old_index )r!r r=r'zgroupsize mismatch: got z , local has r!) isinstancelistr!rstackr/r*shapearanger+reshape) local_permsrCrr!lplocalGSbaseperm2ds &&& r compose_final_permrVws+t$$Q&& kZkUU& UCkZ`abUZZ0 ;;DA >P5i[ QCPP> LL<< 09 < B B1a HDlKNN%,,ejjNQ RF >>" [s)D3cdV^8dQhR\PR\R\P/#)rperm total_columnsr)rrr)rs"r r r s) * * *S *U\\ *r c\VP44pW!8dV#\P!VVVPVP R7p\P !W3^R7#)zEAppend identity-mapped tail columns left outside the full GAR groups.rr=)rr(rrMr r!cat)rXrYcoveredtails&& r extend_perm_with_tailr^sV$**,G <<jj{{  D 99d\q ))r c\P!V4p\P!VP4VPR7W&V#)z Compute the inverse of a permutation vector. Args: perm (Tensor): A 1D tensor containing a permutation of indices. Returns: inv (Tensor): The inverse permutation such that inv[perm] == torch.arange(len(perm)). rG)r empty_likerMr(r!)rXinvs& r invert_permrbs5   4 C TZZ\$++>CI Jr cNV^8dQh/^\9d\R,;R&#)rNr)__conditional_annotations__r)rs"r r r s ('TD[' r )r<) rdrgptqmodel.utilsrrrrr6rDrVr^rbr )rds@r rfs[ (n#'' ! !J) /3 ))X: * r