+ "ݳj@ 6^RIHt^RItRRRlltRRltR#)) annotationsNc$V^8dQhRRRRRR/#)masktorch.Tensor | Noneseq_lenz int | Nonereturn)formats"Z/home/luca/screen-checking/.venv/lib/python3.14/site-packages/gptqmodel/utils/attn_mask.py __annotate__r s#,","0,":,"Qd,"cVfR#TpVP\P8wd+\P!V^84'dV^8pMV^8pVP 4^8XdVP ^4^8XdkVP ^4^8Xd VR,pEMVP ^4VP ^48XdVR,P^R7pMVR,pMVP 4^8Xd!VP ^4^8Xd VR,pMVP 4^8XdM}VeYVP 4^8dDVP R 4V8Xd.VP VP ^4R 4RRV13,pM!\R\VP4 24hVP\PR7#) a* Normalize a variety of HF attention mask formats to a boolean keep-mask [B, S]. True = keep (attended), False = drop (padding/fully-masked). Accepts typical HF forms: - [B, S] with 1/0 - [B, 1, 1, S] 'extended' masks with {0 or positive} keep and {negative large} masked - [B, 1, S] (rare) Ndim.z"Unsupported attention_mask shape: )dtype)NNNrrr)rr)rrr) rtorchboolanyrsizereshape ValueErrortupleshapeto)rrms&& r normalize_seq_maskr sb | Aww%** 99QU  aAQA uuw!|q Q 66!9>* A VVAY!&&) #$  "A* A A!&&)q. gJ A   1557Q;166":3H !&&)R(hwh7AA% BSATUV V 44ejj4 !!rc$V^8dQhRRRRRR/#)rxz torch.Tensor keep_mask_bsrr r )r s"r r r :s"44,46I4l4rc HVeVP4^8dV#VP^4VP^4r2VPW#38wd3\RVP RW#3 R\ VP4 24h\ V4Uu.uFq@WAV,3,NK pp\ V4^8XdAVPRVPR,,VPVPR7#\P!V^R7P4#uupi)z Apply [B, S] keep-mask to a tensor x of shape [B, S, ...]. Returns a flattened tensor of shape [N_kept, ...] (collapses batch/time on the kept rows). If keep_mask is None or x doesn't have [B, S, ...] leading dims, returns x unchanged. z Mask shape z does not match leading dims z of x=:rNN)rdevicer)r) rrrAssertionErrorrrangelen new_zerosrr$rcat contiguous)r!r"BSb kept_rowss&& r apply_keep_mask_btr/:s quuw{ 66!9affQiqaV#{<+=+=*>>[]^\b[ccijopqpwpwjxiyz{{16a9111o%&&I9 9~{{4!''"+-QWWQXX{NN 99YA & 1 1 33:sD)N) __future__rrrr/r rr r1s # ,"^4r