GitHub

blackwell_fa_ws_pipelined.py

nargs["desc_o"].block_shape = [BLOCK_M_SPLIT, HEAD_DIM] configs = [ # triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'NUM_BUFFERS_KV': 3, 'NUM_BUFFERS_QK': 1, 'NUM ...