o (i@sddlmZddlZddlZddlmZ  ddejdejdejdedeejd eejd dfd d Z   ddejd ejdejdedeejdeejdeejd dfddZ dS))OptionalN)rms_norm_kernelxweightoutepsin_scale out_scalereturnc Cs|j\}}t|}tdtd|d} t|fdid|d|d|d|dd |d d d dd |d|d|dd|d|d|d|d ud|d uddddd| d S)z[RMS norm. Computes `out[i,j] = x[i,j] * weight[j] / sqrt(eps + sum(x[i]^2) / n)`.  nbx_ptrx_strider x_scale_ptrr_ptrNr_stridew_ptro_ptro_stride o_scale_ptrEPS BLOCK_SIZE HAS_IN_SCALE HAS_OUT_SCALE HAS_OUTPUTT HAS_RESIDUALF num_warps)shapetritonnext_power_of_2maxminrstride) rrrrrr rr block_sizerr r S/home/ubuntu/veenaModal/venv/lib/python3.10/site-packages/flashinfer/triton/norm.pyrms_norm sP       r)residualx_out x_in_scale x_out_scalec Cs>|j\}}|j|jks J|d|dksJt|} tdt| d} t|fdid|d|d|d|dd|d|d |dd |d |d |d urX|dn$dd|d|d| d|d ud|d ud|d uddd| d Sd|d|d| d|d ud|d ud|d uddd| d S)zmIn-place RMS norm with fused residual addition. Computes `r = r + x`, followed by `x = rmsnorm(r)`. rr rrrrrrrrrrNrrrrrrrTrr )r!r&r"r#r%cdivr) rr*rrr+r,r-rrr'rr r r(rms_norm_add_residual1sx             r/)NN)NNN) typingrtorchr"flashinfer.triton.kernels.normrTensorfloatr)r/r r r r(sR   -