+ j 6^RIt^RIHtRRltRRRlltR#)N)PagedAttentionCachecdV^8dQhR\PR\R\P/#) hidden_statesn_repreturn)torchTensorint)formats"e/home/luca/screen-checking/.venv/lib/python3.14/site-packages/transformers/integrations/sdpa_paged.py __annotate__rs. U UU\\ U# U%,, UcVPwr#rEV^8XdV#VR,PW#WV4pVPW#V,WE4#)z This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) )NNNrNrr)shapeexpandreshape)rrbatchnum_key_value_headsslenhead_dims&& r repeat_kvrsU 2?1D1D.E z!"23::5W\dlmM  e(CT TTrc:V^8dQhR\PPR\PR\PR\PR\PR,R\R\R,R \ \PR3,/#) rmodulequerykeyvalueattention_maskNdropoutscalingr)r nnModuler floattuple)r s"r rrs55 HHOO5 <<5 5 << 5 LL4' 5  5T\5 5<< 5rc VPRR4pVf \R4hVPVVVPVR,VR,R7wr#VP ^^4P ^4pVP ^^4P ^4p\ VR4'd+\W P4p\W0P4pTp VP4pVP4pVP4p\PPPVVVV VVRR7p V P ^^4P4p V R3#) cacheNz`paged|sdpa` was called without a paged attention cache. This implementation expects the packed inputs and the 4D mask that continuous batching prepares; on a standard forward it would attend bidirectionally. Use `sdpa` for a standard forward. read_index write_index) key_states value_states layer_idxr(r)num_key_value_groupsF) attn_mask dropout_pscale is_causal)pop ValueErrorupdater, transpose unsqueezehasattrrr- contiguousr r" functionalscaled_dot_product_attention) rrrrrr r!kwargsr' causal_mask attn_outputs &&&&&&&, r sdpa_attention_paged_forwardr>sX)/ 7D(AE } B  "",'=) JC --1  ' ' *C OOAq ! + +A .Ev-..889%!rr rAs F U55r