@@ -825,17 +825,11 @@ void gpt2_backward(GPT2 *model) {
825825
826826 // get the pointers of the weights for this layer
827827 float * l_ln1w = params .ln1w + l * C ;
828- float * l_ln1b = params .ln1b + l * C ;
829828 float * l_qkvw = params .qkvw + l * 3 * C * C ;
830- float * l_qkvb = params .qkvb + l * 3 * C ;
831829 float * l_attprojw = params .attprojw + l * C * C ;
832- float * l_attprojb = params .attprojb + l * C ;
833830 float * l_ln2w = params .ln2w + l * C ;
834- float * l_ln2b = params .ln2b + l * C ;
835831 float * l_fcw = params .fcw + l * 4 * C * C ;
836- float * l_fcb = params .fcb + l * 4 * C ;
837832 float * l_fcprojw = params .fcprojw + l * C * 4 * C ;
838- float * l_fcprojb = params .fcprojb + l * C ;
839833 // get the pointers of the gradients of the weights for this layer
840834 float * dl_ln1w = grads .ln1w + l * C ;
841835 float * dl_ln1b = grads .ln1b + l * C ;
@@ -855,17 +849,13 @@ void gpt2_backward(GPT2 *model) {
855849 float * l_ln1_rstd = acts .ln1_rstd + l * B * T ;
856850 float * l_qkv = acts .qkv + l * B * T * 3 * C ;
857851 float * l_atty = acts .atty + l * B * T * C ;
858- float * l_preatt = acts .preatt + l * B * NH * T * T ;
859852 float * l_att = acts .att + l * B * NH * T * T ;
860- float * l_attproj = acts .attproj + l * B * T * C ;
861853 float * l_residual2 = acts .residual2 + l * B * T * C ;
862854 float * l_ln2 = acts .ln2 + l * B * T * C ;
863855 float * l_ln2_mean = acts .ln2_mean + l * B * T ;
864856 float * l_ln2_rstd = acts .ln2_rstd + l * B * T ;
865857 float * l_fch = acts .fch + l * B * T * 4 * C ;
866858 float * l_fch_gelu = acts .fch_gelu + l * B * T * 4 * C ;
867- float * l_fcproj = acts .fcproj + l * B * T * C ;
868- float * l_residual3 = acts .residual3 + l * B * T * C ;
869859 // get the pointers of the gradients of the activations for this layer
870860 float * dl_ln1 = grads_acts .ln1 + l * B * T * C ;
871861 float * dl_qkv = grads_acts .qkv + l * B * T * 3 * C ;
0 commit comments