Back to Flashmla

README

csrc/sm100/decode/head128/README.md

latest185 B
Original Source

Head128 decoding kernels are located at csrc/sm100/prefill/sparse/fwd_for_small_topk/head128/instantiations/phase1_decode_k512.cu (for k_dim = 512) or simulated using 2x head64 kernel