Skip to content

vulkan: use mul_mat_vec_id for high-expert-count MoE decode #25862

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
essentialols wants to merge 1 commit into
base: master
Choose a base branch
from
+13 −1
Draft

vulkan: use mul_mat_vec_id for high-expert-count MoE decode #25862

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
14 changes: 13 additions & 1 deletion ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10279,7 +10279,19 @@ static bool ggml_vk_use_mul_mat_vec_id(const struct ggml_cgraph * cgraph, int no
ggml_tensor * dst = cgraph->nodes[node_idx];
ggml_tensor * src0 = dst->src[0];
ggml_tensor * src2 = dst->src[2];
return (src2->ne[1] <= 8) && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type));
const bool type_ok = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type));
// Small batches (single-token decode) always use the vec path.
if (src2->ne[1] <= 8 && type_ok) {
return true;
}
// High-expert-count MoE (>64 experts, e.g. DeepSeek-V3/V4-class 256-expert models):
// route the decode regime (batch <= 256) to mul_mat_vec_id. On RADV the mul_mat_id
// shader for >=256 experts stalls during pipeline compilation at high -ngl; the vec
// path avoids this and is also faster per-dispatch for token generation.
if (src0->ne[2] > 64 && src2->ne[1] <= 256 && type_ok) {
return true;
}
return false;
}

static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx) {
Expand Down