Commit 15758862 authored by rocking's avatar rocking
Browse files

Fix stride for splitkv kernel

parent e148767a
...@@ -686,7 +686,7 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -686,7 +686,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
const ck_tile::index_t batch_stride_o_acc = (nhead * max_seqlen_q * hdim_v); const ck_tile::index_t batch_stride_o_acc = (nhead * max_seqlen_q * hdim_v);
const ck_tile::index_t batch_stride_o = (nhead * shape_seqlen_q * hdim_v); const ck_tile::index_t batch_stride_o = (nhead * shape_seqlen_q * hdim_v);
// setup split_stride_* arguments (only used in split-kv kernel) // setup split_stride_* arguments (only used in split-kv kernel)
const ck_tile::index_t split_stride_lse_acc = (batch * nhead * shape_seqlen_q); const ck_tile::index_t split_stride_lse_acc = (shape_batch * nhead * shape_seqlen_q);
const ck_tile::index_t split_stride_o_acc = (batch * nhead * max_seqlen_q * hdim_v); const ck_tile::index_t split_stride_o_acc = (batch * nhead * max_seqlen_q * hdim_v);
return fmha_fwd_args{q_buf.GetDeviceBuffer(), return fmha_fwd_args{q_buf.GetDeviceBuffer(),
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment