Skip to content

Check if ffn_up and ffn_gate are of the same type before using fmoe #495

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

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Check if ffn_up and ffn_gate are of the same type before using fmoe #495

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
fa54b58
Check if ffn_up and ffn_gate are of the same type before using fmoe
Jun 6, 2025
fdfad30
Just leave the check.
Jun 6, 2025
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
22 changes: 21 additions & 1 deletion src/llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3447,6 +3447,26 @@ static bool llama_kv_cache_init(
buft_layer_count[llama_default_buffer_type_cpu(true)] = n_layer;
}

//if (cparams.fused_moe_up_gate) {
// int nbad = 0;
// for (int i = 0; i < (int) n_layer; i++) {
// auto& layer = model.layers[i];
// if (layer.ffn_gate_exps && layer.ffn_up_exps && layer.ffn_gate_exps->type != layer.ffn_up_exps->type) {
// ++nbad;
// }
// }
// if (nbad > 0) {
// if (nbad == (int)n_layer) {
// LLAMA_LOG_WARN("=============== ffn_up and ffn_gate are of different type => disabling fmoe\n");
// const_cast<llama_cparams&>(cparams).fused_moe_up_gate = false;
// }
// else {
// LLAMA_LOG_WARN("=============== ffn_up and ffn_gate are of different in %d out of %d layers, where fmoe will be disabled\n",
// nbad, (int)n_layer);
// }
// }
//}

// create a context for each buffer type
std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
for (auto & it : buft_layer_count) {
Expand Down Expand Up @@ -9841,7 +9861,7 @@ llm_expert_gating_func_type gating_op,
}

ggml_tensor * par;
if (lctx.cparams.fused_moe_up_gate) {
if (lctx.cparams.fused_moe_up_gate && up_exps->type == gate_exps->type) {
par = ggml_moe_up_gate(ctx, up_exps, gate_exps, cur, selected_experts, type_op == LLM_FFN_SILU ? GGML_UNARY_OP_SILU : GGML_UNARY_OP_GELU);
} else {
ggml_tensor * up = llm_build_lora_mm_id(lctx, ctx, up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
Expand Down