Skip to content

ggml : adjust logic for offloading ops to weight's backend #25832

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
ggerganov merged 2 commits into from
Jul 27, 2026
+34 −20
Merged

ggml : adjust logic for offloading ops to weight's backend #25832

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9291c45
ggml : adjust logic for offloading ops to weight's backend
ggerganov Jul 17, 2026
9f364c7
llama : dsv4 graph fixes
ggerganov Jul 17, 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
43 changes: 26 additions & 17 deletions ggml/src/ggml-backend.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -906,26 +906,35 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
}

// operations with weights are preferably run on the same backend as the weights
for (int i = 0; i < GGML_MAX_SRC; i++) {
const struct ggml_tensor * src = tensor->src[i];
if (src == NULL) {
continue;
}
// skip ROPE since the rope freqs tensor is too small to choose a backend based on it
// not an ideal solution
if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
// check if a backend with higher prio wants to offload the op
if (sched->op_offload && src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
for (int b = 0; b < src_backend_id; b++) {
if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
SET_CAUSE(tensor, "1.off");
return b;
// TODO: there are exceptions (see below) - not an ideal solution
bool allow = true;

// skip ROPE since the rope freqs tensor is too small to choose a backend based on it
allow = allow && tensor->op != GGML_OP_ROPE;

// skip FLASH_ATTN_EXT since the sinks tensor is too small to choose a based based on it
allow = allow && tensor->op != GGML_OP_FLASH_ATTN_EXT;

@eapache eapache Jul 17, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think this fixes the issue I reported, but probably leaves an equivalent issue around if flash-attention is disabled with -fa off? The sinks tensor will get attached somewhere else (a soft_max?) and cause the same issue there.

@ggerganov ggerganov Jul 17, 2026

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, it will likely cause the soft max ops to be offloaded when -fa off because that's where the sinks tensor will go. It's not a big problem - this logic has to be reworked either way.


if (allow) {
for (int i = 0; i < GGML_MAX_SRC; i++) {
const struct ggml_tensor * src = tensor->src[i];
if (src == NULL) {
continue;
}
if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
// check if a backend with higher prio wants to offload the op
if (sched->op_offload && src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
for (int b = 0; b < src_backend_id; b++) {
if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
SET_CAUSE(tensor, "1.off");
return b;
}
}
}
SET_CAUSE(tensor, "1.wgt%d", i);
return src_backend_id;
}
SET_CAUSE(tensor, "1.wgt%d", i);
return src_backend_id;
}
}

Expand Down
5 changes: 3 additions & 2 deletions src/llama-context.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2472,11 +2472,12 @@ llm_graph_cb llama_context::graph_get_cb() const {
ggml_set_name(cur, name);
}

// norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
// - norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
// - force the last op of the layer on the specified backend to avoid running it on the backend of the next layer due to scheduling
// FIXME: fix in ggml_backend_sched
const bool full_offload = model.n_gpu_layers() > model.hparams.n_layer_all;
if (ubatch.n_tokens < 32 || full_offload) {
if (il != -1 && strcmp(name, "norm") == 0) {
if (il != -1 && (strcmp(name, "norm") == 0 || strcmp(name, "l_last") == 0)) {
const auto & dev_layer = model.dev_layer(il);
for (const auto & backend : backends) {
if (ggml_backend_get_device(backend.get()) == dev_layer) {
Expand Down
6 changes: 5 additions & 1 deletion src/models/deepseek4.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1133,6 +1133,10 @@ llama_model_deepseek4::graph::graph(const llama_model & model, const llm_graph_p
&post, &comb, il);
cb(cur, "hc_ffn_pre", il);

ggml_build_forward_expand(gf, residual);
ggml_build_forward_expand(gf, post);
ggml_build_forward_expand(gf, comb);

cur = build_norm(cur, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);

Expand Down Expand Up @@ -1175,7 +1179,7 @@ llama_model_deepseek4::graph::graph(const llama_model & model, const llm_graph_p

inpL = build_hc_post(cur, residual, post, comb, il);
inpL = build_cvec(inpL, il);
cb(inpL, "l_out", il);
cb(inpL, "l_last", il);
}

if (inp_out_ids) {
Expand Down
Loading