Skip to content

ggml-cuda: add dequant-float matvec (mmvdq) for Q4_K/Q5_K/Q6_K on RDN… #61

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
jimw567 merged 4 commits into from
Jul 26, 2026
+783 −3
Merged

ggml-cuda: add dequant-float matvec (mmvdq) for Q4_K/Q5_K/Q6_K on RDN… #61

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
e2624bb
ggml-cuda: add dequant-float matvec (mmvdq) for Q4_K/Q5_K/Q6_K on RDN…
Jul 19, 2026
488afd1
GGML_CUDA_DQ_ROWS env parsing hardened, Row bounds guard
Jul 22, 2026
fe04f5c
Remove the guard break
Jul 25, 2026
5f1585f
Reduce vgpr number
Jul 26, 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
44 changes: 41 additions & 3 deletions ggml/src/ggml-cuda/ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,6 +36,7 @@
#include "ggml-cuda/mmq.cuh"
#include "ggml-cuda/mmvf.cuh"
#include "ggml-cuda/mmvq.cuh"
#include "ggml-cuda/mmvdq.cuh"
#include "ggml-cuda/norm.cuh"
#include "ggml-cuda/opt-step-adamw.cuh"
#include "ggml-cuda/opt-step-sgd.cuh"
Expand Down Expand Up @@ -1790,6 +1791,22 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
ggml_cuda_mul_mat_f(ctx, src0, src1, nullptr, dst);
return;
}
// GGML_CUDA_DQ_Q6K (unset = arch default, 0 = off, non-zero = on).
const bool dq_default = GGML_CUDA_CC_IS_RDNA3_5(cc);
if (ggml_cuda_dq_mmv_enabled(dq_default) && ne11 == 1
&& (src0->type == GGML_TYPE_Q4_K || src0->type == GGML_TYPE_Q5_K
|| (src0->type == GGML_TYPE_Q6_K && ggml_cuda_dq_q6k_enabled(dq_default)))
&& ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)
&& ne02 == 1 && ne03 == 1 && ne12 == 1 && ne13 == 1 && src0->ne[0] % QK_K == 0) {
if (src0->type == GGML_TYPE_Q4_K) {
ggml_cuda_mul_mat_vec_dq_q4_K(ctx, src0, src1, dst);
} else if (src0->type == GGML_TYPE_Q5_K) {
ggml_cuda_mul_mat_vec_dq_q5_K(ctx, src0, src1, dst);
} else {
ggml_cuda_mul_mat_vec_dq_q6_K(ctx, src0, src1, dst);
}
return;
}
if (ggml_cuda_should_use_mmvq(src0->type, cc, ne11)) {
ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
return;
Expand Down Expand Up @@ -3578,6 +3595,23 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
}

if (ggml_cuda_should_fuse_mul_mat_vec_q(up)) {
const bool dq_default = GGML_CUDA_CC_IS_RDNA3_5(ggml_cuda_info().devices[cuda_ctx->device].cc);
if (ggml_cuda_dq_mmv_enabled(dq_default) && ids == nullptr
&& ggml_get_glu_op(glu) == GGML_GLU_OP_SWIGLU
&& (src0->type == GGML_TYPE_Q4_K || src0->type == GGML_TYPE_Q5_K
|| (src0->type == GGML_TYPE_Q6_K && ggml_cuda_dq_q6k_enabled(dq_default)))
&& gate->src[0]->type == src0->type && ggml_are_same_shape(src0, gate->src[0])
&& src1->type == GGML_TYPE_F32 && glu->type == GGML_TYPE_F32
&& src1->ne[1] == 1 && src0->ne[0] % QK_K == 0
&& ggml_is_contiguous(src0) && ggml_is_contiguous(gate->src[0])
&& ggml_is_contiguous(src1) && ggml_is_contiguous(glu)
&& src0->ne[2] == 1 && src0->ne[3] == 1 && src1->ne[2] == 1 && src1->ne[3] == 1) {
ggml_cuda_mul_mat_vec_dq_glu(*cuda_ctx, src0, gate->src[0], src1, glu);
fused_mul_mat_vec = true;
fused_node_count = 3;
break;
}

ggml_cuda_mm_fusion_args_host fusion_data{};
fusion_data.gate = gate->src[0];
fusion_data.glu_op = ggml_get_glu_op(glu);
Expand Down Expand Up @@ -4124,9 +4158,13 @@ static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph
GGML_UNUSED(cgraph);
#endif

static bool enable_graph_optimization = [] {
const char * env = getenv("GGML_CUDA_GRAPH_OPT");
return env != nullptr && atoi(env) == 1;
static bool enable_graph_optimization = [cuda_ctx] {
const char * env = getenv("GGML_CUDA_GRAPH_OPT");
if (env != nullptr) {
return atoi(env) == 1;
}
const int cc = ggml_cuda_info().devices[cuda_ctx->device].cc;
return GGML_CUDA_CC_IS_RDNA3_5(cc);
}();

if (!enable_graph_optimization) {
Expand Down
Loading
Loading