CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama-bench: add --n-cpu-moe support (#15952)

Browse Source 
* llama-bench: add --n-cpu-moe support

Support --n-cpu-moe in llama-bench the same way it is supported by
llama-server.
This commit is contained in:
jacekpoplawski
2025-09-16 16:17:08 +02:00
committed by GitHub
parent 77475530b8
commit 8ff206097c
3 changed files with 90 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
common/arg.cpp
View File

@@ -2548,7 +2548,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--cpu-moe", "-cmoe"},
"keep all Mixture of Experts (MoE) weights in the CPU",
[](common_params & params) {
params.tensor_buft_overrides.push_back({"\\.ffn_(up|down|gate)_exps", ggml_backend_cpu_buffer_type()});
params.tensor_buft_overrides.push_back(llm_ffn_exps_cpu_override());
}
).set_env("LLAMA_ARG_CPU_MOE"));
add_opt(common_arg(
@@ -2561,7 +2561,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
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(string_format("blk\\.%d\\.ffn_(up|down|gate)_exps", i));
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()});
}
}
@@ -2570,7 +2570,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--cpu-moe-draft", "-cmoed"},
"keep all Mixture of Experts (MoE) weights in the CPU for the draft model",
[](common_params & params) {
params.speculative.tensor_buft_overrides.push_back({"\\.ffn_(up|down|gate)_exps", ggml_backend_cpu_buffer_type()});
params.speculative.tensor_buft_overrides.push_back(llm_ffn_exps_cpu_override());
}
).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CPU_MOE_DRAFT"));
add_opt(common_arg(
@@ -2582,7 +2582,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
for (int i = 0; i < value; ++i) {
static std::list<std::string> buft_overrides_draft;
buft_overrides_draft.push_back(string_format("blk\\.%d\\.ffn_(up|down|gate)_exps", i));
buft_overrides_draft.push_back(llm_ffn_exps_block_regex(i));
params.speculative.tensor_buft_overrides.push_back({buft_overrides_draft.back().c_str(), ggml_backend_cpu_buffer_type()});
}
}

14
common/common.h
View File

@@ -734,6 +734,20 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
}
//
// MoE utils
//
const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate)_exps";
static std::string llm_ffn_exps_block_regex(int idx) {
return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);
}
static llama_model_tensor_buft_override llm_ffn_exps_cpu_override() {
return { LLM_FFN_EXPS_REGEX, ggml_backend_cpu_buffer_type() };
}
//
// training utils
//