Skip to content

Various fused ops around expert selection #840

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

Merged
merged 9 commits into from
Oct 19, 2025
Merged

Various fused ops around expert selection #840

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
8f5f93e
Fuse sigmoid+add+grouped_topk+get_rows (CPU)
Oct 18, 2025
2c66dc8
Fix CPU + CUDA
Oct 18, 2025
f3ff1a5
Minor
Oct 19, 2025
8fe2bb9
Fuse sigmoid+add+topk+get_rows (CUDA)
Oct 19, 2025
18d9f4f
Fuse sigmoid+add+topk+get_rows (CPU)
Oct 19, 2025
c8ed454
Fuse topk+view+get_rows+reshape+softmax (CPU)
Oct 19, 2025
b79aad9
Fuse topk+view+get_rows+reshape+softmax (CUDA)
Oct 19, 2025
0fb9d49
cpu: turn off the openai topk fusing for now
Oct 19, 2025
1d70b89
Also fuse sum_rows and div
Oct 19, 2025
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
42 changes: 39 additions & 3 deletions ggml/src/ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3173,7 +3173,25 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
ggml_cuda_op_relu(ctx, dst);
break;
case GGML_UNARY_OP_SIGMOID:
ggml_cuda_op_sigmoid(ctx, dst);
if (i + 5 < cgraph->n_nodes &&
cgraph->nodes[i+1]->op == GGML_OP_RESHAPE &&
cgraph->nodes[i+2]->op == GGML_OP_ADD &&
cgraph->nodes[i+3]->op == GGML_OP_ARGSORT &&
cgraph->nodes[i+4]->op == GGML_OP_VIEW &&
cgraph->nodes[i+5]->op == GGML_OP_GET_ROWS) {
cuda_glm45moe_experts(ctx, cgraph->nodes[i+5], cgraph->nodes[i+4]);
i += 5;
}
else if (i + 4 < cgraph->n_nodes &&
cgraph->nodes[i+1]->op == GGML_OP_RESHAPE &&
cgraph->nodes[i+2]->op == GGML_OP_ADD &&
cgraph->nodes[i+3]->op == GGML_OP_GROUPED_TOPK &&
cgraph->nodes[i+4]->op == GGML_OP_GET_ROWS) {
cuda_bailingmoev2_experts(ctx, cgraph->nodes[i+4], cgraph->nodes[i+4]);
i += 4;
} else {
ggml_cuda_op_sigmoid(ctx, dst);
}
break;
case GGML_UNARY_OP_HARDSIGMOID:
ggml_cuda_op_hardsigmoid(ctx, dst);
Expand Down Expand Up @@ -3315,10 +3333,28 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
ggml_cuda_op_pool2d(ctx, dst);
break;
case GGML_OP_SUM_ROWS:
ggml_cuda_op_sum_rows(ctx, dst);
if (i + 1 < cgraph->n_nodes &&
cgraph->nodes[i+1]->op == GGML_OP_DIV &&
cgraph->nodes[i+1]->src[1] == dst &&
cgraph->nodes[i+1]->src[0] == dst->src[0]) {
ggml_cuda_op_sum_rows_div(ctx, cgraph->nodes[i+1]);
++i;
} else {
ggml_cuda_op_sum_rows(ctx, dst);
}
break;
case GGML_OP_ARGSORT:
ggml_cuda_op_argsort(ctx, dst);
if (i + 5 < cgraph->n_nodes &&
cgraph->nodes[i+1]->op == GGML_OP_VIEW &&
cgraph->nodes[i+2]->op == GGML_OP_GET_ROWS &&
cgraph->nodes[i+3]->op == GGML_OP_RESHAPE &&
cgraph->nodes[i+4]->op == GGML_OP_SOFT_MAX &&
cgraph->nodes[i+5]->op == GGML_OP_RESHAPE) {
cuda_openai_experts(ctx, dst, cgraph->nodes[i+4]);
i += 5;
} else {
ggml_cuda_op_argsort(ctx, dst);
}
break;
case GGML_OP_ARGSORT_THRESH:
ggml_cuda_op_argsort_thresh(ctx, dst);
Expand Down
Loading