j dZddlZddlmZddlmZmZmZ ddlm Z n#e e e f$rdZ YnwxYwejejfZdZdZd ZejeedS) aProjections that MMGP row views compute in place from a fused weight interleaving its projections head by head (mmgp.row_views): the Triton GEMM reads each projection's rows where they are (as precise as cuBLAS or better: FP32 accumulation, one rounding); the PyTorch path runs one cuBLAS product per head. Registered with MMGP at import, which then keeps such weights fused instead of copying their projections apart at load.N) row_views)choosecuda_inferencetriton_dtype_supported) rowmap_tritonctduoht|tdoQ|j|jkoA|jo:|ddko!|ddko|dzdkS)NT)dtypeshipr@r)_tritonr_DTYPESdtypeis_cudastride)xweighthead_dims DC:\Users\johnn\OneDrive\Documents\AI\shared\kernels\rowmap_linear.py_triton_usablers 4  AN1W$$O$O$O A ' A,2N A?E}}Q?O?OST?T AYZYaYabcYdYdhiYi Anvy{n{@AoABc tdt|||rtj||||||||S||jd||z f}t |||D]} | |z|z|z||zz} |dd| |z | |z |zf} |(tj||| | |zj | Ttj || | |z||| | |zj | |S)zColumns [start, stop) (multiples of head_dim) of projection `slot` of a weight whose rows are (heads, slots, head_dim), for x of shape (tokens, in_features).zInterleaved Projections) owned_tritonrN)out) rrr rows new_emptyshaperangetorchmmTaddmm) rrbiasslotslotsrstartstoprcolumnrowtargets rrrs6'nQPX6Y6YZZZQ|AvtT5(E4PPP ++qwqz4%<0 1 1CtX..__ 5(83dXoEQQQv~'@@@A < HQs3>124& A A A A A KSx/0!VCht tjdtj|S)zWhether MMGP should keep a fused interleaved weight of this dtype: the Triton GEMM makes its projections as fast as split ones.Nr rcuda)r rrr, is_availablerdevicecurrent_device)rrs r availabler0'st 4  aEW$4 aB!9K aPUPZPgPgPiPi a&u|FEJr<s$$ DDDDDDDDDD*******Wl+GGG >5= )BBB    bbb  $ *****s  ,,