Skip to content

cuda: opt-in MoE epilogue fusions (GGML_CUDA_MOE_EP_FUSION) #423

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
bri-prism merged 1 commit into from
Oct 31, 2026
Merged

cuda: opt-in MoE epilogue fusions (GGML_CUDA_MOE_EP_FUSION) #423

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
122 changes: 121 additions & 1 deletion ggml/src/ggml-cuda/ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4341,6 +4341,65 @@ static bool ggml_cuda_match_moe_weighted_reduction(
}


// GGML_CUDA_MOE_EP_VERIFY=1: compare against the existing fused reduction (REPEAT/GET_ROWS result input);
// =2: compare against the fully unfused MUL, MUL, VIEW, ADD nodes
static int ggml_cuda_moe_ep_verify_mode() {
static const int m = getenv("GGML_CUDA_MOE_EP_VERIFY") ? atoi(getenv("GGML_CUDA_MOE_EP_VERIFY")) : 0;
return m;
}

// GGML_CUDA_MOE_EP_FUSION bit mask: 1 = scale gather into the weighted reduction, 2 = router RMS_NORM+SCALE+MUL, 4 = router matvec + top-k
static int ggml_cuda_moe_ep_bits() {
static const int bits = getenv("GGML_CUDA_MOE_EP_FUSION") ? atoi(getenv("GGML_CUDA_MOE_EP_FUSION")) : 0;
return bits;
}
static bool ggml_cuda_moe_ep_fusion_enabled() {
return (ggml_cuda_moe_ep_bits() & 1) != 0;
}

// REPEAT(w_s [1,n_expert,1] -> [1,n_expert,n_tokens]) -> GET_ROWS(.., ids) -> MUL(experts, .) -> weighted reduction.
// The gather and the repeat only feed the expert-scale MUL, so the reduction kernel can read w_s[ids] directly.
static bool ggml_cuda_match_moe_scale_gather(
const ggml_cgraph * cgraph,
int node_idx,
ggml_cuda_moe_weighted_reduction_match & match,
const ggml_tensor *& table,
const ggml_tensor *& ids) {
if (node_idx + 3 >= cgraph->n_nodes) {
return false;
}
const ggml_tensor * rep = cgraph->nodes[node_idx];
const ggml_tensor * get = cgraph->nodes[node_idx + 1];
if (rep->op != GGML_OP_REPEAT || get->op != GGML_OP_GET_ROWS || get->src[0] != rep ||
rep->type != GGML_TYPE_F32 || get->type != GGML_TYPE_F32) {
return false;
}
const ggml_tensor * w_s = rep->src[0];
ids = get->src[1];
if (!w_s || !ids || w_s->type != GGML_TYPE_F32 || !ggml_is_contiguous(w_s) || w_s->ne[0] != 1 || w_s->ne[2] != 1 ||
w_s->ne[3] != 1 || ids->type != GGML_TYPE_I32 || ids->ne[2] != 1 || ids->ne[3] != 1 ||
ids->nb[0] != sizeof(int32_t) || ids->nb[1] % sizeof(int32_t) != 0 ||
rep->ne[0] != 1 || rep->ne[1] != w_s->ne[1] || rep->ne[3] != 1 ||
get->ne[0] != 1 || get->ne[1] != ids->ne[0] || get->ne[2] != ids->ne[1] || get->ne[3] != 1 ||
rep->ne[2] != ids->ne[1] || !ggml_is_contiguous(rep) || !ggml_is_contiguous(get)) {
return false;
}
if (!ggml_cuda_match_moe_weighted_reduction(cgraph, node_idx + 2, match) || match.expert_scale != get) {
return false;
}
const int count = match.node_count + 2;
std::vector<ggml_op> ops(count);
for (int j = 0; j < count; ++j) {
ops[j] = cgraph->nodes[node_idx + j]->op;
}
const int output_idx = node_idx + count - 1;
if (!ggml_can_fuse_subgraph(cgraph, node_idx, count, ops.data(), &output_idx, 1)) {
return false;
}
table = w_s;
return true;
}

static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph,
int node_idx,
std::initializer_list<enum ggml_op> ops,
Expand Down Expand Up @@ -5395,13 +5454,51 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
}
}

if (node->op == GGML_OP_MUL) {
// GGML_CUDA_MOE_EP_FUSION=1: fold the per-expert scale gather (REPEAT + GET_ROWS of down_exps_s by the selected ids)
// into the weighted reduction kernel
if (node->op == GGML_OP_REPEAT && ggml_cuda_moe_ep_fusion_enabled()) {
ggml_cuda_moe_weighted_reduction_match match;
const ggml_tensor * table = nullptr;
const ggml_tensor * ids = nullptr;
if (ggml_cuda_match_moe_scale_gather(cgraph, i, match, table, ids)) {
const int count = match.node_count + 2;
const int output_idx = i + count - 1;
if (ggml_cuda_check_fusion_memory_ranges(cgraph, i, count, &output_idx, 1)) {
// GGML_CUDA_MOE_EP_VERIFY=1: run the fused kernel into scratch, run the unfused nodes, compare bit-for-bit at the output
if (ggml_cuda_moe_ep_verify_mode() != 0) {
static float * scratch = nullptr; static int tag = 0;
if (!scratch) { CUDA_CHECK(cudaMalloc(&scratch, (size_t) 32*1024*1024)); }
const size_t need = ggml_nelements(match.dst)*sizeof(float);
if (need <= (size_t) 32*1024*1024 && tag < 1000000) {
ggml_tensor tmp = *match.dst;
tmp.data = scratch;
ggml_cuda_op_moe_weighted_reduction(*cuda_ctx, match.experts, nullptr, match.weights, &tmp, table, ids);
ggml_cuda_tail_verify_pending()[match.dst] = { scratch, tag++ };
}
return 0;
}
ggml_cuda_op_moe_weighted_reduction(
*cuda_ctx, match.experts, nullptr, match.weights, match.dst, table, ids);
return count - 1;
}
}
}

if (node->op == GGML_OP_MUL && !(ggml_cuda_moe_ep_verify_mode() == 2 && ggml_cuda_moe_ep_fusion_enabled())) {
ggml_cuda_moe_weighted_reduction_match match;
if (ggml_cuda_match_moe_weighted_reduction(cgraph, i, match)) {
const int output_idx = i + match.node_count - 1;
if (ggml_cuda_check_fusion_memory_ranges(cgraph, i, match.node_count, &output_idx, 1)) {
ggml_cuda_op_moe_weighted_reduction(
*cuda_ctx, match.experts, match.expert_scale, match.weights, match.dst);
if (!ggml_cuda_tail_verify_pending().empty()) {
auto it = ggml_cuda_tail_verify_pending().find(match.dst);
if (it != ggml_cuda_tail_verify_pending().end()) {
ggml_cuda_op_tail_verify_cmp(*cuda_ctx, match.dst, it->second.scratch, it->second.tag);
static int nl = 0;
if ((++nl & (nl - 1)) == 0) fprintf(stderr, "TAILVERIFY compare launches captured/issued: %d\n", nl);
}
}
return match.node_count - 1;
}
}
Expand Down Expand Up @@ -6162,6 +6259,25 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
return 1;
}

const int ep_out2 = i + 2;
if ((ggml_cuda_moe_ep_bits() & 2) && i + 2 < cgraph->n_nodes && cgraph->nodes[i + 1]->op == GGML_OP_SCALE &&
cgraph->nodes[i + 1]->src[0] == node && cgraph->nodes[i + 2]->op == GGML_OP_MUL &&
(cgraph->nodes[i + 2]->src[0] == cgraph->nodes[i + 1] || cgraph->nodes[i + 2]->src[1] == cgraph->nodes[i + 1]) &&
ggml_can_fuse_subgraph(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_SCALE, GGML_OP_MUL }, { i + 2 }) &&
ggml_cuda_check_fusion_memory_ranges(cgraph, i, 3, &ep_out2, 1)) {
if (ggml_cuda_moe_ep_verify_mode() != 0) {
static float * scratch = nullptr; static int tag = 0;
if (!scratch) { CUDA_CHECK(cudaMalloc(&scratch, (size_t) 32*1024*1024)); }
ggml_tensor * mul_n = cgraph->nodes[i + 2];
if (ggml_nelements(mul_n)*sizeof(float) <= (size_t) 32*1024*1024 && tag < 1000000 &&
ggml_cuda_op_rms_norm_scale_mul_fused(*cuda_ctx, node, cgraph->nodes[i + 1], mul_n, scratch)) {
ggml_cuda_tail_verify_pending()[mul_n] = { scratch, tag++ };
}
} else if (ggml_cuda_op_rms_norm_scale_mul_fused(*cuda_ctx, node, cgraph->nodes[i + 1], cgraph->nodes[i + 2])) {
return 2;
}
}

if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_SCALE }, {})) {
ggml_cuda_op_rms_norm_scale_fused(*cuda_ctx, node, cgraph->nodes[i + 1]);
return 1;
Expand Down Expand Up @@ -6820,6 +6936,10 @@ static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph
if (match.expert_scale != nullptr) {
params->add_alloc_dep(
params->user_data, const_cast<ggml_tensor *>(match.expert_scale), match.dst);
if (ggml_cuda_moe_ep_fusion_enabled() && match.expert_scale->op == GGML_OP_GET_ROWS && match.expert_scale->src[1]) {
params->add_alloc_dep(
params->user_data, const_cast<ggml_tensor *>(match.expert_scale->src[1]), match.dst);
}
}
i += match.node_count - 1;
}
Expand Down
29 changes: 23 additions & 6 deletions ggml/src/ggml-cuda/moe-weighted-reduction.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,20 +5,27 @@ static __global__ void moe_weighted_reduction_f32(const float * __restrict__ exp
const float * __restrict__ weights,
float * __restrict__ dst,
const int64_t n_embd,
const int n_expert_used) {
const int n_expert_used,
const float * __restrict__ scale_table,
const int32_t * __restrict__ scale_ids,
const int64_t ids_s1) {
const int64_t token = blockIdx.x;
const int64_t col = (int64_t) blockIdx.y * blockDim.x + threadIdx.x;
if (col >= n_embd) {
return;
}

const uint64_t first_row = (uint64_t) token * n_expert_used;
const float first_scale = expert_scale != nullptr ? expert_scale[first_row] : 1.0f;
// scale_table != nullptr: the per-expert scale is gathered here, scale_table[ids[token, expert]], instead of
// being read from the REPEAT + GET_ROWS result (same float values)
const float first_scale = scale_table != nullptr ? scale_table[scale_ids[token * ids_s1]] :
expert_scale != nullptr ? expert_scale[first_row] : 1.0f;
float sum = (experts[first_row * n_embd + col] * first_scale) * weights[first_row];

for (int expert = 1; expert < n_expert_used; ++expert) {
const uint64_t row = first_row + expert;
const float scale = expert_scale != nullptr ? expert_scale[row] : 1.0f;
const float scale = scale_table != nullptr ? scale_table[scale_ids[token * ids_s1 + expert]] :
expert_scale != nullptr ? expert_scale[row] : 1.0f;
sum += (experts[row * n_embd + col] * scale) * weights[row];
}
dst[token * n_embd + col] = sum;
Expand All @@ -31,18 +38,23 @@ static void launch_moe_weighted_reduction(const float * experts,
int64_t n_embd,
int64_t n_tokens,
int n_expert_used,
const float * scale_table,
const int32_t * scale_ids,
int64_t ids_s1,
cudaStream_t stream) {
constexpr int threads = 256;
const dim3 blocks(n_tokens, (n_embd + threads - 1) / threads, 1);
moe_weighted_reduction_f32
<<<blocks, threads, 0, stream>>>(experts, expert_scale, weights, dst, n_embd, n_expert_used);
<<<blocks, threads, 0, stream>>>(experts, expert_scale, weights, dst, n_embd, n_expert_used, scale_table, scale_ids, ids_s1);
}

void ggml_cuda_op_moe_weighted_reduction(ggml_backend_cuda_context & ctx,
const ggml_tensor * experts,
const ggml_tensor * expert_scale,
const ggml_tensor * weights,
ggml_tensor * dst) {
ggml_tensor * dst,
const ggml_tensor * scale_table,
const ggml_tensor * scale_ids) {
GGML_ASSERT(experts->type == GGML_TYPE_F32);
GGML_ASSERT(weights->type == GGML_TYPE_F32);
GGML_ASSERT(expert_scale == nullptr || expert_scale->type == GGML_TYPE_F32);
Expand All @@ -51,6 +63,8 @@ void ggml_cuda_op_moe_weighted_reduction(ggml_backend_cuda_context & ctx,
GGML_ASSERT(ggml_is_contiguous(weights));
GGML_ASSERT(expert_scale == nullptr || ggml_is_contiguous(expert_scale));
GGML_ASSERT(ggml_is_contiguous(dst));
GGML_ASSERT(scale_ids == nullptr || (scale_ids->type == GGML_TYPE_I32 && scale_ids->nb[0] == sizeof(int32_t) &&
scale_ids->nb[1] % sizeof(int32_t) == 0 && scale_table != nullptr && scale_table->type == GGML_TYPE_F32));

const int64_t n_embd = experts->ne[0];
const int64_t n_expert_used = experts->ne[1];
Expand All @@ -60,6 +74,9 @@ void ggml_cuda_op_moe_weighted_reduction(ggml_backend_cuda_context & ctx,
launch_moe_weighted_reduction((const float *) experts->data,
expert_scale ? (const float *) expert_scale->data : nullptr,
(const float *) weights->data,
(float *) dst->data, n_embd, n_tokens, (int) n_expert_used, stream);
(float *) dst->data, n_embd, n_tokens, (int) n_expert_used,
scale_table ? (const float *) scale_table->data : nullptr,
scale_ids ? (const int32_t *) scale_ids->data : nullptr,
scale_ids ? (int64_t) (scale_ids->nb[1] / sizeof(int32_t)) : 0, stream);
CUDA_CHECK(cudaGetLastError());
}
4 changes: 3 additions & 1 deletion ggml/src/ggml-cuda/moe-weighted-reduction.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,4 +4,6 @@ void ggml_cuda_op_moe_weighted_reduction(ggml_backend_cuda_context & ctx,
const ggml_tensor * experts,
const ggml_tensor * expert_scale,
const ggml_tensor * weights,
ggml_tensor * dst);
ggml_tensor * dst,
const ggml_tensor * scale_table = nullptr,
const ggml_tensor * scale_ids = nullptr);
79 changes: 79 additions & 0 deletions ggml/src/ggml-cuda/norm.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -952,6 +952,85 @@ void ggml_cuda_op_add_rms_norm_scale_fused(
}
}

// RMS_NORM -> SCALE -> MUL(weight): same reduction and expression order as the rms_norm_f32 kernel the unfused RMS_NORM takes
// (the scalar path, or with vec the float4 path with one float4 per thread), followed by the elementwise scale and multiply,
// so the result is bit-identical to the three separate nodes.
template <bool vec>
static __global__ void rms_norm_scale_mul_f32(const float * x, const float * w, float * dst, const int ncols,
const int64_t stride_row, const float eps, const float scale_out) {
ggml_cuda_pdl_lc();
extern __shared__ float s_sum[];
const int row = blockIdx.x;
const int tid = threadIdx.x;
ggml_cuda_pdl_sync();
const float * xc = x + row*stride_row;
float * dstc = dst + (int64_t) row*ncols;
if constexpr (vec) {
const int n4 = ncols / 4;
float4 xv = make_float4(0.0f, 0.0f, 0.0f, 0.0f);
float tmp = 0.0f;
if (tid < n4) {
xv = ((const float4 *) xc)[tid];
tmp = __fadd_rn(tmp, rms_norm_sumsq4(xv));
}
tmp = block_reduce<block_reduce_method::SUM, 1024>(tmp, s_sum);
const float scale = rms_norm_row_scale(tmp, ncols, eps);
if (tid < n4) {
const float4 wv = ((const float4 *) w)[tid];
float4 r;
r.x = __fmul_rn(__fmul_rn(scale_out, __fmul_rn(scale, xv.x)), wv.x);
r.y = __fmul_rn(__fmul_rn(scale_out, __fmul_rn(scale, xv.y)), wv.y);
r.z = __fmul_rn(__fmul_rn(scale_out, __fmul_rn(scale, xv.z)), wv.z);
r.w = __fmul_rn(__fmul_rn(scale_out, __fmul_rn(scale, xv.w)), wv.w);
((float4 *) dstc)[tid] = r;
}
} else {
float tmp = 0.0f;
for (int col = tid; col < ncols; col += 1024) {
const float xi = xc[col];
tmp += xi * xi;
}
tmp = block_reduce<block_reduce_method::SUM, 1024>(tmp, s_sum);
const float mean = tmp / ncols;
const float scale = rsqrtf(mean + eps);
for (int col = tid; col < ncols; col += 1024) {
dstc[col] = (scale_out * (scale * xc[col])) * w[col];
}
}
}

bool ggml_cuda_op_rms_norm_scale_mul_fused(ggml_backend_cuda_context & ctx, ggml_tensor * rms, ggml_tensor * scale_t, ggml_tensor * mul_t, float * verify_scratch) {
const ggml_tensor * src0 = rms->src[0];
const ggml_tensor * w = mul_t->src[0] == scale_t ? mul_t->src[1] : mul_t->src[0];
if (src0->type != GGML_TYPE_F32 || w->type != GGML_TYPE_F32 || mul_t->type != GGML_TYPE_F32 || scale_t->type != GGML_TYPE_F32 ||
src0->ne[0] < 1024 || src0->ne[2] != 1 || src0->ne[3] != 1 || src0->nb[0] != sizeof(float) ||
!ggml_is_contiguous(w) || ggml_nelements(w) != src0->ne[0] || !ggml_is_contiguous(mul_t) ||
!ggml_are_same_shape(mul_t, src0)) {
return false;
}
float eps, scale;
memcpy(&eps, rms->op_params, sizeof(float));
memcpy(&scale, (const float *) scale_t->op_params + 0, sizeof(float));
if (!(eps >= 0.0f) || ((const float *) scale_t->op_params)[1] != 0.0f) {
return false;
}
// The unfused RMS_NORM takes the float4 path when its operands are 16-byte aligned and ncols/4 <= 1024 fits one float4
// per thread; wider rows use a different float4 geometry that this kernel does not reproduce, so they are not fused.
const int ncols = (int) src0->ne[0];
const bool vec = rms_norm_vector_enabled() && ncols % 4 == 0 && ggml_cuda_is_aligned(src0, 16) &&
((uintptr_t) rms->data & 15) == 0;
if (vec && ncols / 4 > 1024) {
return false;
}
if (vec && (!ggml_cuda_is_aligned(w, 16) || ((uintptr_t) scale_t->data & 15) != 0 || ((uintptr_t) mul_t->data & 15) != 0)) {
return false;
}
const ggml_cuda_kernel_launch_params lp{ dim3(src0->ne[1], 1, 1), dim3(1024, 1, 1), 32 * sizeof(float), ctx.stream() };
ggml_cuda_kernel_launch(vec ? rms_norm_scale_mul_f32<true> : rms_norm_scale_mul_f32<false>, lp, (const float *) src0->data, (const float *) w->data, verify_scratch ? verify_scratch : (float *) mul_t->data,
(int) src0->ne[0], (int64_t) (src0->nb[1] / sizeof(float)), eps, scale);
return true;
}

void ggml_cuda_op_rms_norm_scale_fused(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * scale_tensor) {
const ggml_tensor * src0 = dst->src[0];
float * dst_d = (float *) scale_tensor->data;
Expand Down
1 change: 1 addition & 0 deletions ggml/src/ggml-cuda/norm.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,6 +21,7 @@ void ggml_cuda_op_add_rms_norm_scale_fused(

void ggml_cuda_op_rms_norm_fused(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * mul_tensor);

bool ggml_cuda_op_rms_norm_scale_mul_fused(ggml_backend_cuda_context & ctx, ggml_tensor * rms, ggml_tensor * scale_t, ggml_tensor * mul_t, float * verify_scratch = nullptr);
void ggml_cuda_op_rms_norm_scale_fused(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * scale_tensor);

void ggml_cuda_op_rms_norm_fused_add(ggml_backend_cuda_context & ctx,
Expand Down
Loading