Skip to content

perf(ds4): enable sparse gate-up MMQ prefill #673

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
davide221 merged 2 commits into from
Aug 28, 2026
Merged

perf(ds4): enable sparse gate-up MMQ prefill #673

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
29d2181
perf(ds4): enable sparse gate-up MMQ prefill
dusterbloom Aug 27, 2026
465a85b
fix(cuda): honor force cublas for DS4 mix pair
dusterbloom Aug 27, 2026
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 server/deps/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2816,7 +2816,19 @@ static bool ggml_cuda_try_fuse_mul_mat_glu(
}

const int64_t ncols = ids ? src1->ne[2] : src1->ne[1];
if (ggml_cuda_should_use_mmq(
#ifdef GGML_CUDA_FORCE_CUBLAS
// The global force-cuBLAS policy is authoritative over this RDNA 3.5 default.
constexpr bool default_ds4_mix_gate_up_mmq = false;
#else
const char * mix_mmq = std::getenv("DFLASH_DS4_MIX_MMQ_PREFILL");
const bool default_ds4_mix_gate_up_mmq =
src0->type == GGML_TYPE_Q2_1_ROCMFP2_MIX &&
ids != nullptr &&
GGML_CUDA_CC_IS_RDNA3_5(cc) &&
(!mix_mmq || !(mix_mmq[0] == '0' && mix_mmq[1] == '\0'));
#endif
if (default_ds4_mix_gate_up_mmq ||
Comment thread
cubic-dev-ai[bot] marked this conversation as resolved.
ggml_cuda_should_use_mmq(
src0->type, cc, ncols,
ids ? src0->ne[2] : /*n_experts=*/0)) {
ggml_cuda_mul_mat_q_pair(
Expand Down
Loading