+ j >^RIt^RIHt^RIHtRRltRRltR#)N)nn)PagedAttentionCachecdV^8dQhR\PR\R\P/#) hidden_statesn_repreturn)torchTensorint)formats"f/home/luca/screen-checking/.venv/lib/python3.14/site-packages/transformers/integrations/eager_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\PR\PR\PR\PR\PR,R\/#)rmodulequerykeyvalueattention_maskNscaling)rModuler r float)r s"rrrs]=%=% II=% <<=% =% << =% LL4' =%  =%rc ~VPRR4pVf \R4hVPVVVPVR,VR,R7wr#VP ^^4P ^4pVP ^^4P ^4p\ VR4'd+\W P4p\W0P4p\V\4'd&\VR^4pV^8XgVfRMR p WI,p MTp \P!WP ^^44V,p V e W,p \ VR 4'dVPP^R^^4P!VP"^,RVP"R,R4p \P$!W.RR 7p WP'RR R 7P(, p \*P,P/V R\P0R7P3VP44p V RRR13,p MI\*P,P/V R\P0R7P3VP44p \P!W4p V P ^^4P74p W3#)cacheNz`paged|eager` 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 `eager` for a standard forward. read_index write_index) key_states value_states layer_idxr&r'num_key_value_groupssliding_windowfull_attentionsliding_attentionsinks)dimT)r0keepdim)r0dtype.)pop ValueErrorupdater* transpose unsqueezehasattrrr+ isinstancedictgetattrr matmulr/rrrcatmaxvaluesr functionalsoftmaxfloat32tor2 contiguous)rrrrr r!kwargsr%r, layer_type causal_mask attn_weightsr/ attn_outputs&&&&&&, reager_paged_attention_forwardrLsY)/ 7D(AE } C  "",'=) JC --1  ' ' *C OOAq ! + +A .Ev-..889%!rOs F U=%r