Skip to content

CUDA: gated_delta_net - share per-token k/q via shared memory for long prefills #54

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
roberteg16 wants to merge 1 commit into
base: gfx11
Choose a base branch
from
+451 −0
Draft

CUDA: gated_delta_net - share per-token k/q via shared memory for long prefills #54

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
26 changes: 26 additions & 0 deletions ggml/src/ggml-cuda/gated_delta_net.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,5 @@
#include "gated_delta_net.cuh"
#include "gated_delta_net_chunked.cuh"
#include "ggml-cuda/common.cuh"

template <int S_v, bool KDA, bool keep_rs_t>
Expand Down Expand Up @@ -294,6 +295,31 @@ static void ggml_cuda_op_gated_delta_net_impl(
state_slot_stride = cache->slot_stride;
}

if (ggml_cuda_gdn_chunked_supported(kda, keep_rs, S_v, n_tokens)) {
ggml_cuda_gdn_chunked_args args = {};
args.q = q_d;
args.k = k_d;
args.v = v_d;
args.g = g_d;
args.beta = b_d;
args.state_in = s_d;
args.dst = dst_d;
args.state_out = state_d;
args.S_v = S_v;
args.H = H;
args.n_tokens = n_tokens;
args.n_seqs = n_seqs;
args.sq1 = sq1; args.sq2 = sq2; args.sq3 = sq3;
args.sv1 = sv1; args.sv2 = sv2; args.sv3 = sv3;
args.sb1 = sb1; args.sb2 = sb2; args.sb3 = sb3;
args.neqk1 = neqk1;
args.rq3 = rq3;
args.scale = scale;

ggml_cuda_gdn_chunked(ctx, args);
return;
}

if (kda) {
if (keep_rs) {
launch_gated_delta_net<true, true>(q_d, k_d, v_d, g_d, b_d, s_d, dst_d, state_d,
Expand Down
Loading
Loading