tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-07-29 05:33:37 -04:00
Code Issues Packages Projects Releases Wiki Activity

metal : fuse add, mul + add tests (#14596)

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-07-18 20:37:26 +03:00
committed by GitHub
parent 9fb1042ce6
commit bf9087f59a
8 changed files with 578 additions and 172 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
src/llama-graph.cpp
View File

@@ -907,20 +907,25 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
cb(cur, "ffn_moe_weighted", il);
}
ggml_tensor * cur_experts[LLAMA_MAX_EXPERTS] = { nullptr };
assert(n_expert_used > 0);
// order the views before the adds
for (uint32_t i = 0; i < hparams.n_expert_used; ++i) {
cur_experts[i] = ggml_view_2d(ctx0, experts, n_embd, n_tokens, experts->nb[2], i*experts->nb[1]);
ggml_build_forward_expand(gf, cur_experts[i]);
}
// aggregate experts
// note: here we explicitly use hparams.n_expert_used instead of n_expert_used
// to avoid potentially a large number of add nodes during warmup
// ref: https://github.com/ggml-org/llama.cpp/pull/14753
ggml_tensor * moe_out = nullptr;
for (uint32_t i = 0; i < hparams.n_expert_used; ++i) {
ggml_tensor * cur_expert = ggml_view_2d(ctx0, experts, n_embd, n_tokens,
experts->nb[2], i*experts->nb[1]);
ggml_tensor * moe_out = cur_experts[0];
if (i == 0) {
moe_out = cur_expert;
} else {
moe_out = ggml_add(ctx0, moe_out, cur_expert);
}
for (uint32_t i = 1; i < hparams.n_expert_used; ++i) {
moe_out = ggml_add(ctx0, moe_out, cur_experts[i]);
}
if (hparams.n_expert_used == 1) {