o hz @sddlZddlmZmZddlmZddlmZddZeddd ied d d ied ejfd d Z eddd ied dd ied ejfddZ Gdddej j Z e jZdS)N) heuristicsjit)languagenext_power_of_2cCs|dkrdS|dkr dSdS)Nii )Nr r N/var/www/html/ai/venv/lib/python3.10/site-packages/triton/ops/cross_entropy.py num_warpss rcC t|dSNr rnargsr r r  rBLOCKcCrrrrr r r rrc Cstd}td|}t||}||||}||||} ||||} tj|||ktd d} | tj} | t| d} tt t | d| } tj | | ||kdt t| } t ||| dSNrinf)maskother)r) tl program_idarangeloadfloattofloat32maxlogsumexpstore debug_barrier) LOGITSPROBSIDXLOSSr rrowcolsidx WRIT_PROBS READ_PROBSlogitsprobsr r r _forwards    r3cCrrrrr r r r)rcCrrrrr r r r*rc Cstd}td|}t||}||||}tj|||ktdd }t|tj}||k} t||} || | } tj|| |j j ||kddSr) rrrrrr%r r!r&dtype element_ty) r)r*DPROBSr rr,r-r.r2deltadoutdinr r r _backward)s   "r:c@s$eZdZeddZeddZdS)_cross_entropyc s~|jtjks Jdjj}}jdtj|||d}tj||d}fdd}t||||||||S)Nz(Indices are expected to be of type long.)r4devicecsfSNnumeloptr1n_colsr r rHz(_cross_entropy.forward..)r4torchint64r=shape empty_liker3save_for_backward) clsctxr1indicesr=r4result neg_logprobsgridr rCr forward>s  z_cross_entropy.forwardcs<|j\}jdfdd}t|||dfS)aWe know d(-log(p[i])/dlogit[k] = -id_mat[i,k] + p[k] so we initialize the gradient as neg_logprobs, so we can just exponentiate to get p[k], which is most of what we need... neg_logprobs will be modified in place to become the gradient we want r<csfSr>r?rArDrOr r rZrEz)_cross_entropy.backward..N) saved_tensorsrHr:)rKrL dneg_logprobsrMrPr rRr backwardNs  z_cross_entropy.backwardN)__name__ __module__ __qualname__ classmethodrQrUr r r r r;=s  r;)rFrrrrrr constexprr3r:autogradFunctionr;apply cross_entropyr r r r s   "