in benchmark/bench_flash_mla.py [0:0]
def _mla_attn_kernel(
Q_nope,
Q_pe,
Kv_c_cache,
K_pe_cache,
Req_to_tokens,
B_seq_len,
O,
sm_scale,
stride_q_nope_bs,
stride_q_nope_h,
stride_q_pe_bs,
stride_q_pe_h,
stride_kv_c_bs,
stride_k_pe_bs,
stride_req_to_tokens_bs,
stride_o_b,
stride_o_h,
stride_o_s,
BLOCK_H: tl.constexpr,
BLOCK_N: tl.constexpr,
NUM_KV_SPLITS: tl.constexpr,
PAGE_SIZE: tl.constexpr,
HEAD_DIM_CKV: tl.constexpr,
HEAD_DIM_KPE: tl.constexpr,