ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-01 03:59:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fused FFN_UP+FFN_GATE op (#741)

Browse Source 
* Fused up+gate+unary for regular (not MoE) FFN - CPU

* WIP CUDA

* Seems to be working on CUDA

For a dense model we get 2-3% speedup for PP and ~0.6% for TG.

* Add command line option

This time the option is ON by default, and one needs to turn it
off via -no-fug or --no-fused-up-gate

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-08-31 18:16:36 +03:00
committed by GitHub
parent f22a9ef95a
commit b66cecca45
10 changed files with 276 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
common/common.cpp
View File

@@ -1004,6 +1004,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
params.fused_moe_up_gate = true;
return true;
}
if (arg == "-no-fug" || arg == "--no-fused-up-gate") {
params.fused_up_gate = false;
return true;
}
if (arg == "-ser" || arg == "--smart-expert-reduction") {
CHECK_ARG
auto values = string_split_pairs<int,float>(argv[i], ',');
@@ -1760,6 +1764,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
options.push_back({ "*", "-mla, --mla-use", "enable MLA (default: %d)", params.mla_attn });
options.push_back({ "*", "-amb, --attention-max-batch", "max batch size for attention computations (default: %d)", params.attn_max_batch});
options.push_back({ "*", "-fmoe, --fused-moe", "enable fused MoE (default: %s)", params.fused_moe_up_gate ? "enabled" : "disabled" });
options.push_back({ "*", "-no-fug, --no-fused-up-gate", "disaable fused up-gate (default: %s)", params.fused_up_gate ? "enabled" : "disabled" });
options.push_back({ "*", "-ser, --smart-expert-reduction,","experts reduction (default: %d,%g)", params.min_experts, params.thresh_experts});
options.push_back({ "*", "-p, --prompt PROMPT", "prompt to start generation with\n"
"in conversation mode, this will be used as system prompt\n"
@@ -2660,6 +2665,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
cparams.mla_attn = params.mla_attn;
cparams.attn_max_batch = params.attn_max_batch;
cparams.fused_moe_up_gate = params.fused_moe_up_gate;
cparams.fused_up_gate = params.fused_up_gate;
cparams.min_experts = params.min_experts;
cparams.thresh_experts = params.thresh_experts;
@@ -3756,6 +3762,7 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
fprintf(stream, "mla_attn: %d # default: 0\n", params.mla_attn);
fprintf(stream, "attn_max_batch: %d # default: 0\n", params.attn_max_batch);
fprintf(stream, "fused_moe: %s # default: false\n", params.fused_moe_up_gate ? "true" : "false");
fprintf(stream, "fused_up_gate: %s # default: true\n", params.fused_up_gate ? "true" : "false");
fprintf(stream, "ser: %d,%g # defaulr: -1,0\n", params.min_experts, params.thresh_experts);
fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);