Skip to content

A slightly better version of offloading only activated experts #762

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
merged 1 commit into from
Sep 5, 2025
Merged

A slightly better version of offloading only activated experts #762

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
79 changes: 47 additions & 32 deletions ggml/src/ggml-backend.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1857,6 +1857,10 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
struct ggml_backend_sched_split * splits = sched->splits;

std::vector<int32_t> ids;
std::vector<uint32_t> unique_ids;
ggml_tensor * last_ids_tensor = nullptr;


for (int i = 0; i < sched->n_splits; i++) {
#if IK_PRINT_TIMING
Expand All @@ -1865,14 +1869,9 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
struct ggml_backend_sched_split * split = &splits[i];
int split_backend_id = split->backend_id;
ggml_backend_t split_backend = sched->backends[split_backend_id];
ggml_backend_t last_input_backend = nullptr;

int cur_arg = 0;
std::vector<int32_t> ids;
std::set<int32_t> unique_ids;

//printf("Graph split %d has %d inputs:\n", i, split->n_inputs);
//for (int j = 0; j < split->n_inputs; j++) printf(" %s, %s\n", split->inputs[j]->name,
// split->inputs[j]->src[0] ? split->inputs[j]->src[0]->name : "none");

// copy the input tensors to the split backend
for (int j = 0; j < split->n_inputs; j++) {
Expand Down Expand Up @@ -1903,32 +1902,46 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
node->src[cur_arg] == input_cpy &&
(node->op == GGML_OP_MUL_MAT_ID || node->op == GGML_OP_MOE_FUSED_UP_GATE)) {

if (ids.empty()) {
// find the ids
ggml_tensor * ids_tensor = node->op == GGML_OP_MUL_MAT_ID ? node->src[2] : node->src[3];
ids.resize(ggml_nbytes(ids_tensor) / sizeof(int32_t));
if (input_backend != last_input_backend) {
ggml_backend_synchronize(input_backend);
last_input_backend = input_backend;
}

ggml_backend_tensor_get_async(split_backend, ids_tensor, ids.data(), 0, ggml_nbytes(ids_tensor));
//printf("node: %s have %d inputs, processing input %d\n", node->name, split->n_inputs, j);
ggml_tensor * ids_tensor = node->op == GGML_OP_MUL_MAT_ID ? node->src[2] : node->src[3];
auto ids_backend = split_backend;

ggml_backend_synchronize(split_backend);
// if the ids tensor is also an input of the split, it may not have been copied yet to the split backend
// in that case, we use the original ids tensor
for (int jj = j + 1; jj < split->n_inputs; ++jj) {
if (ids_tensor == tensor_copy(split->inputs[jj], split_backend_id, sched->cur_copy)) {
ids_tensor = split->inputs[jj];
ids_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[jj]);
break;
}
}

int n_expert = input->ne[2];

if (ids_tensor != last_ids_tensor) {
ids.resize(ggml_nbytes(ids_tensor) / sizeof(int32_t));

ggml_backend_tensor_get_async(ids_backend, ids_tensor, ids.data(), 0, ggml_nbytes(ids_tensor));

ggml_backend_synchronize(ids_backend);

unique_ids.resize((n_expert + 31)/32);
std::memset(unique_ids.data(), 0, unique_ids.size()*sizeof(uint32_t));
for (int64_t i1 = 0; i1 < ids_tensor->ne[1]; i1++) {
for (int64_t i0 = 0; i0 < ids_tensor->ne[0]; i0++) {
int32_t id = ids[i1 * ids_tensor->nb[1]/sizeof(int32_t) + i0 * ids_tensor->nb[0]/sizeof(int32_t)];
unique_ids.insert(id);
unique_ids[id >> 5] |= (1u << (id & 31));
}
}

// group consecutive experts and copy them together
GGML_ASSERT(!unique_ids.empty());

last_ids_tensor = ids_tensor;
}

auto it = unique_ids.begin();
int32_t first_id = *it;
int32_t last_id = first_id;

auto copy_experts = [&](int32_t first_id, int32_t last_id) {
const size_t expert_size = (node->op == GGML_OP_MUL_MAT_ID || node->op == GGML_OP_MOE_FUSED_UP_GATE) ? input->nb[2] : input->nb[1];
const size_t expert_offset = first_id * expert_size;
Expand All @@ -1944,20 +1957,22 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s

};

for (++it; it != unique_ids.end(); ++it) {
const int32_t id = *it;

if (id == last_id + 1) {
last_id = id;
continue;
}

copy_experts(first_id, last_id);
auto next_on_id = [&unique_ids, n_expert] (int id) {
while (id < n_expert && (unique_ids[id >> 5] & (1u << (id & 31))) == 0) ++id;
return id;
};
auto next_off_id = [&unique_ids, n_expert] (int id) {
while (id < n_expert && (unique_ids[id >> 5] & (1u << (id & 31))) != 0) ++id;
return id;
};

first_id = id;
last_id = id;
int first_id = next_on_id(0);
while (first_id < n_expert) {
int last_id = next_off_id(first_id+1);
copy_experts(first_id, last_id-1);
first_id = next_on_id(last_id);
}
copy_experts(first_id, last_id);

if (node->op == GGML_OP_MOE_FUSED_UP_GATE) ++cur_arg;
} else
// try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
Expand Down