Skip to content

llama : add --n-cpu-ffn option #26622

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
CISC merged 3 commits into from
Aug 27, 2026
Merged

llama : add --n-cpu-ffn option #26622

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
0b35214
common : dedupe --n-cpu-moe / --spec-draft-n-cpu-moe override loops
John-194 Aug 5, 2026
b1c214d
common : add --n-cpu-ffn to CPU-offload dense FFN weights of first N …
John-194 Aug 5, 2026
0345d3a
common : generalize llm_ffn_block_regex over the FFN regex, drop TODO
John-194 Aug 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
25 changes: 14 additions & 11 deletions common/arg.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2671,14 +2671,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
if (value < 0) {
throw std::invalid_argument("invalid value");
}
for (int i = 0; i < value; ++i) {
// keep strings alive and avoid leaking memory by storing them in a static vector
static std::list<std::string> buft_overrides;
buft_overrides.push_back(llm_ffn_exps_block_regex(i));
params.tensor_buft_overrides.push_back({buft_overrides.back().c_str(), ggml_backend_cpu_buffer_type()});
}
llm_add_n_cpu_ffn_overrides(value, LLM_FFN_EXPS_REGEX, params.tensor_buft_overrides);
}
).set_env("LLAMA_ARG_N_CPU_MOE"));
// TODO: Optimize. Current logic touches layers from 0 blindly. Optimize by prioritizing layers with the largest FFNs first, thus reducing the total offloaded layer count.
Comment thread
John-194 marked this conversation as resolved.
Resolved
add_opt(common_arg(
{"-ncffn", "--n-cpu-ffn"}, "N",
"keep the dense FFN weights of the first N layers in the CPU\n"
"(dense models; for MoE expert weights use --n-cpu-moe)",
[](common_params & params, int value) {
if (value < 0) {
throw std::invalid_argument("invalid value");
}
llm_add_n_cpu_ffn_overrides(value, LLM_FFN_DENSE_REGEX, params.tensor_buft_overrides);
}
).set_env("LLAMA_ARG_N_CPU_FFN"));
GGML_ASSERT(params.n_gpu_layers < 0); // string_format would need to be extended for a default >= 0
add_opt(common_arg(
{"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
Expand Down Expand Up @@ -3982,11 +3989,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
if (value < 0) {
throw std::invalid_argument("invalid value");
}
for (int i = 0; i < value; ++i) {
static std::list<std::string> buft_overrides_draft;
buft_overrides_draft.push_back(llm_ffn_exps_block_regex(i));
params.speculative.draft.tensor_buft_overrides.push_back({buft_overrides_draft.back().c_str(), ggml_backend_cpu_buffer_type()});
}
llm_add_n_cpu_ffn_overrides(value, LLM_FFN_EXPS_REGEX, params.speculative.draft.tensor_buft_overrides);
}
).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_N_CPU_MOE"));

Expand Down
14 changes: 13 additions & 1 deletion common/common.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,7 @@
#include "ggml.h"
#include "llama.h"

#include <list>
#include <set>
#include <sstream>
#include <string>
Expand Down Expand Up @@ -1083,11 +1084,13 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
}

//
// MoE utils
// FFN offload utils
//

const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate|gate_up)_(ch|)exps";

const char * const LLM_FFN_DENSE_REGEX = "\\.ffn_(up|down|gate)\\.";

inline std::string llm_ffn_exps_block_regex(int idx) {
return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);
}

@CISC CISC Aug 26, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

Either adapt this to take the regex and rename or replace it in llama-bench.

@John-194 John-194 Aug 26, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

Done. Should --n-cpu-ffn also become part of llama-bench, mirroring --n-cpu-moe?

@CISC CISC Aug 27, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

It would be useful, doesn't have to be in this PR though.

Expand All @@ -1096,6 +1099,15 @@ inline llama_model_tensor_buft_override llm_ffn_exps_cpu_override() {
return { LLM_FFN_EXPS_REGEX, ggml_backend_cpu_buffer_type() };
}

inline void llm_add_n_cpu_ffn_overrides(int n, const char * ffn_regex, std::vector<llama_model_tensor_buft_override> & overrides) {
// keep strings alive and avoid leaking memory by storing them in a static list
static std::list<std::string> buft_override_strings;
for (int i = 0; i < n; ++i) {
buft_override_strings.push_back(string_format("blk\\.%d%s", i, ffn_regex));
overrides.push_back({buft_override_strings.back().c_str(), ggml_backend_cpu_buffer_type()});
}
}

//
// training utils
//
Expand Down
Loading