+ QV-jRt^RIHtHtHt]!4'd^RIt]P !]4tRt Rt Rt RRlt R Rlt R#) z2HQQ (Half-Quadratic Quantization) integration file)is_hqq_availableis_torch_availableloggingNcDVP4F wrWnK R#N) named_modulesname)modelrmodules& n/Volumes/fast/ai/experiments/ui-tars-smoke/.venv/lib/python3.14/site-packages/transformers/integrations/hqq.pyautoname_modulesr s++-  .cRPVPR4Uu.uF%qR9gK VP4'dK#VNK' up4#uupi).)r layers)joinsplit isnumeric)rns& r name_to_linear_tagr sA 88 3l1>Q5QQ\]\g\g\iQQl mmlsAAAc\4'd^RIHp\4pVP 4FMwr4\ V\ PPX34'gK3VP\V44KO \V4#)) HQQLinear) rhqq.core.quantizersetr isinstancetorchnnLinearaddrlist)r r linear_tagsrr s& r get_linear_tagsr"%sc/%K++-  fuxx : ; ; OO.t4 5.  r cVP4EF,wrEVf.pVPV4\V\PP 4'd\ VP4pWa9dlW,ebW,VPV,n \V4VPV,n VPV,PR4RpRFp\WWR4K \\VP!444^8d\#VVVR7wrVP%R4EK/ W3#)NFT patch_paramshas_been_replaced)W_qmeta)named_childrenappendrrrrrr_modules quant_configtype source_clsrequires_grad_setattrlenr children_prepare_for_hqq_linearpop) r r%r&current_key_namerr linear_tagatt_s &&&& r r4r40s,,.   #! % fehhoo . .+FKK8J)+78D8PENN4(56:6lENN4(3NN4(77> $ 'T*' tFOO%& '! +#:)"3$ A R ?/B  ##r cpaVf.MTp\V4\V4oVPpVPp\ \ S4\ V4, \ V4, 4o\ ;QJdV3RlV4F 'gK RM RM!V3RlV44'd(\PS4pVPV4M\PSV4p\WVR7wrRVRVPRV/VPn V'g\PR4V#) a Prepares nn.Linear layers for HQQ quantization. Since each layer type can have separate quantization parameters, we need to do the following: 1- tag each module with its name via autoname_modules() 2- Extract linear_tags (e.g. ['self_attn.q_proj', ...]) 3- Map quantization parameters as a dictionary linear_tag -> quant_params as HQQLinear expects it, this is referred to as patch_params c3,<"TF qS9xK R#5ir).0keyr!s& r )prepare_for_hqq_linear..ks 6#+ sTFr$r- quant_method skip_moduleszYYZK s 6 6sss 6 666}}[1 L)}}[,? 6 rQsI5AA   H % n "$J,r