Skip to content

Enable split mode graph for on-the-fly merged up/gate experts #1413

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
ikawrakow merged 3 commits into from
Mar 13, 2026
+86 −11
Merged

Enable split mode graph for on-the-fly merged up/gate experts #1413

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f719756
Split mode graph for on-the-fly merged ffn_up/gate_exps
ikawrakow Mar 12, 2026
ec4693b
Cleanup
ikawrakow Mar 12, 2026
c046f7f
Also handle merged bias
ikawrakow Mar 12, 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
81 changes: 79 additions & 2 deletions ggml/src/ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -854,6 +854,85 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_init_tensor([[maybe_unused]
}

GGML_CALL static void ggml_backend_cuda_split_buffer_set_tensor([[maybe_unused]] ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
if (!tensor->extra && tensor->view_src && tensor->view_src->extra) {
// OK, this is an ugly hack, but I don't really see a way to trick the machine into correctly
// loading non-contiguous merged split tensors.
auto view_src = tensor->view_src;
auto extra = (ggml_split_tensor_t *)view_src->extra;
void * extra_ptr;
memcpy(&extra_ptr, view_src->op_params, sizeof(extra_ptr));
if (extra_ptr) {
std::string merged_name = view_src->name;
if (auto pos = merged_name.find("ffn_gate_up_exps.weight"); pos != std::string::npos) {
std::string name = tensor->name;
auto pos_u = name.find("ffn_up_exps.weight");
auto pos_g = name.find("ffn_gate_exps.weight");
if (pos_u != std::string::npos || pos_g != std::string::npos) {
GGML_ASSERT(extra->split_dim == 1);
auto & ranges = *(const std::vector<std::vector<std::pair<int,int>>> *)extra_ptr;
int ne = 0;
for (int is = 0; is < int(ranges.size()); ++is) {
auto & r = ranges[is];
GGML_ASSERT((extra->splits[is] && !r.empty()) || (!extra->splits[is] && r.empty()));
if (r.empty()) continue;
GGML_ASSERT(r.size() == 2);
auto split = extra->splits[is];
ggml_cuda_set_device(is);
int ir = pos_g != std::string::npos ? 0 : 1;
auto p = r[ir];
size_t offset = 0;
if (ir == 1) {
p.first -= tensor->ne[1];
GGML_ASSERT(p.first >= 0);
offset = split->ne[1]/2 * split->nb[1];
}
for (int i02 = 0; i02 < split->ne[2]; ++i02) {
auto dst = (char *)split->data + i02*split->nb[2] + offset;
auto src = (const char *)data + i02*tensor->nb[2] + ne*tensor->nb[1];
CUDA_CHECK(cudaMemcpyAsync(dst, src, p.second*tensor->nb[1], cudaMemcpyHostToDevice, cudaStreamPerThread));
}
ne += p.second;
CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
}
}
return;
}
if (auto pos = merged_name.find("ffn_gate_up_exps.bias"); pos != std::string::npos) {
std::string name = tensor->name;
auto pos_u = name.find("ffn_up_exps.bias");
auto pos_g = name.find("ffn_gate_exps.bias");
if (pos_u != std::string::npos || pos_g != std::string::npos) {
GGML_ASSERT(extra->split_dim == 0);
auto & ranges = *(const std::vector<std::vector<std::pair<int,int>>> *)extra_ptr;
int ne = 0;
for (int is = 0; is < int(ranges.size()); ++is) {
auto & r = ranges[is];
GGML_ASSERT((extra->splits[is] && !r.empty()) || (!extra->splits[is] && r.empty()));
if (r.empty()) continue;
GGML_ASSERT(r.size() == 2);
auto split = extra->splits[is];
ggml_cuda_set_device(is);
int ir = pos_g != std::string::npos ? 0 : 1;
auto p = r[ir];
size_t offset = 0;
if (ir == 1) {
p.first -= tensor->ne[0];
GGML_ASSERT(p.first >= 0);
offset = split->ne[0]/2 * split->nb[0];
}
for (int i01 = 0; i01 < split->ne[1]; ++i01) {
auto dst = (char *)split->data + i01*split->nb[1] + offset;
auto src = (const char *)data + i01*tensor->nb[1] + ne*tensor->nb[0];
CUDA_CHECK(cudaMemcpyAsync(dst, src, p.second*tensor->nb[0], cudaMemcpyHostToDevice, cudaStreamPerThread));
}
ne += p.second;
CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
}
}
return;
}
}
}
if (!tensor->extra) return;
static std::map<ggml_type, int> k_map = {
{ GGML_TYPE_Q4_0_R8 , 8},
Expand Down Expand Up @@ -886,7 +965,6 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_set_tensor([[maybe_unused]]
{ GGML_TYPE_Q8_KV_R8 , 4},
{ GGML_TYPE_Q8_K_R8 , 8},
};
//printf("%s(%s)\n", __func__, tensor->name);

// split tensors must always be set in their entirety at once
GGML_ASSERT(offset == 0);
Expand Down Expand Up @@ -984,7 +1062,6 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_set_tensor([[maybe_unused]]
auto row_size = ggml_row_size(tensor->type, tensor->ne[0]);
std::vector<char> host_buffer;
int ne1 = 0;
int extra_ne1 = 0;
for (int i = 0; i < extra->n_device; ++i) {
auto split = extra->splits[i];
if (!split) continue;
Expand Down
12 changes: 4 additions & 8 deletions src/llama-load-tensors.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3168,6 +3168,9 @@ bool create_tensors_helper::merge_up_gate_exps(const LLM_TN & tn, int i, int bia

layer.ffn_up_gate_exps = ggml_new_tensor_3d(u_ctx, u_meta->type, u_meta->ne[0], u_meta->ne[1] + g_meta->ne[1], u_meta->ne[2]);
snprintf(layer.ffn_up_gate_exps->name, GGML_MAX_NAME, "blk.%d.ffn_gate_up_exps.weight", i);
if (u_ctx == ctx_split) {
split_tensors.insert(layer.ffn_up_gate_exps);
}
layer.ffn_gate_exps = ml.create_tensor_as_view(u_ctx, layer.ffn_up_gate_exps, g_name.c_str(),
{ g_meta->ne[0], g_meta->ne[1], g_meta->ne[2] }, 0);
layer.ffn_up_exps = ml.create_tensor_as_view(u_ctx, layer.ffn_up_gate_exps, u_name.c_str(),
Expand Down Expand Up @@ -3702,13 +3705,6 @@ bool create_tensors_helper::create_tensors() {
LLAMA_LOG_WARN("========================================================\n\n");
ml.merge_qkv = false;
}
if (ml.merge_up_gate_exps && (model.split_mode == LLAMA_SPLIT_MODE_GRAPH || model.split_mode == LLAMA_SPLIT_MODE_ATTN)) {
LLAMA_LOG_WARN("\n========================================================\n");
LLAMA_LOG_WARN("merge_up_gate_exps is not compatible with split mode 'graph'\n");
LLAMA_LOG_WARN(" => turning off merge_up_gate_exps\n");
LLAMA_LOG_WARN("========================================================\n\n");
ml.merge_up_gate_exps = false;
}
switch (model.arch) {
case LLM_ARCH_LLAMA:
case LLM_ARCH_REFACT:
Expand Down Expand Up @@ -4039,7 +4035,7 @@ bool create_tensors_helper::create_tensors() {
prepare_split_tensors(1, ctx_split, layer.ffn_up_gate_exps, layer.split_ffn_up_gate_exps, up_gate_split, mem_used);
prepare_up_gate_split(layer.ffn_up_gate_exps, layer.split_ffn_up_gate_exps);
if (layer.ffn_up_gate_exps_b) {
prepare_split_tensors(1, ctx_split, layer.ffn_up_gate_exps_b, layer.split_ffn_up_gate_exps_b, up_gate_split, mem_used);
prepare_split_tensors(0, ctx_split, layer.ffn_up_gate_exps_b, layer.split_ffn_up_gate_exps_b, up_gate_split, mem_used);
prepare_up_gate_split(layer.ffn_up_gate_exps_b, layer.split_ffn_up_gate_exps_b);
}
} else {
Expand Down
4 changes: 3 additions & 1 deletion src/llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5203,7 +5203,9 @@ struct llama_context * llama_init_from_model(
LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(ctx->sched));
}

llama_repack_up_gate_exps(*ctx);
if (ctx->model.split_mode != LLAMA_SPLIT_MODE_GRAPH) {
llama_repack_up_gate_exps(*ctx);
}

// build worst-case graph
int n_past = cparams.n_ctx - n_tokens;
Expand Down