ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-26 09:09:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

Cleanup

Browse Source
This commit is contained in:
Kawrakow
2026-01-12 15:28:06 +02:00
parent 74dc8aa99e
commit 905bca2e1c
1 changed files with 0 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files

53
src/llama.cpp
View File

@@ -4363,59 +4363,6 @@ static void llama_repack_up_gate_exps(llama_context & lctx) {
}
}
}
/*
ggml_init_params params{lctx.buf_compute_meta.size(), lctx.buf_compute_meta.data(), true};
auto ctx = ggml_init(params);
auto gf = ggml_new_graph_custom(ctx, model.max_nodes(), false);
for (int il = 0; il < int(model.layers.size()); ++il) {
auto & l = model.layers[il];
if (l.ffn_up_gate_exps && l.ffn_up_exps && l.ffn_gate_exps) {
GGML_ASSERT(l.ffn_up_gate_exps->type == l.ffn_up_exps->type && l.ffn_up_gate_exps->type == l.ffn_gate_exps->type);
GGML_ASSERT(l.ffn_up_gate_exps->ne[0] == l.ffn_up_exps->ne[0] && l.ffn_up_gate_exps->ne[0] == l.ffn_gate_exps->ne[0]);
GGML_ASSERT(l.ffn_up_gate_exps->ne[2] == l.ffn_up_exps->ne[2] && l.ffn_up_gate_exps->ne[2] == l.ffn_gate_exps->ne[2]);
GGML_ASSERT(l.ffn_up_gate_exps->ne[1] == l.ffn_up_exps->ne[1] + l.ffn_gate_exps->ne[1]);
printf("%s: repacking up/gate experts in layer %d\n", __func__, il);
auto aux = ggml_dup(ctx, l.ffn_up_exps);
auto ffn_up_gate_exps_flat = ggml_reshape_2d(ctx, l.ffn_up_gate_exps,
l.ffn_up_gate_exps->ne[0], l.ffn_up_gate_exps->ne[1]*l.ffn_up_gate_exps->ne[2]);
auto ffn_up_flat = ggml_reshape_2d(ctx, aux, l.ffn_up_exps->ne[0], l.ffn_up_exps->ne[1]*l.ffn_up_exps->ne[2]);
auto ffn_gate_flat = ggml_reshape_2d(ctx, l.ffn_gate_exps, l.ffn_gate_exps->ne[0], l.ffn_gate_exps->ne[1]*l.ffn_gate_exps->ne[2]);
size_t offset_up_gate = 0;
size_t offset_up = 0;
for (int i2 = 0; i2 < (int)l.ffn_up_gate_exps->ne[2]; ++i2) {
auto dst = ggml_view_2d(ctx, ffn_up_gate_exps_flat, l.ffn_up_exps->ne[0], l.ffn_up_exps->ne[1],
l.ffn_up_exps->nb[1], offset_up_gate);
auto src = ggml_view_2d(ctx, ffn_up_flat, l.ffn_up_exps->ne[0], l.ffn_up_exps->ne[1],
l.ffn_up_exps->nb[1], offset_up);
auto cpy1 = ggml_cpy(ctx, src, dst);
offset_up_gate += l.ffn_up_exps->ne[1]*l.ffn_up_exps->nb[1];
if (i2 < (int)l.ffn_up_gate_exps->ne[2]-1) {
dst = ggml_view_2d(ctx, ffn_up_gate_exps_flat, l.ffn_up_exps->ne[0], l.ffn_up_exps->ne[1],
l.ffn_up_exps->nb[1], offset_up_gate);
src = ggml_view_2d(ctx, ffn_gate_flat, l.ffn_up_exps->ne[0], l.ffn_up_exps->ne[1],
l.ffn_up_exps->nb[1], offset_up);
auto cpy2 = ggml_cpy(ctx, src, dst);
offset_up_gate += l.ffn_up_exps->ne[1]*l.ffn_up_exps->nb[1];
offset_up += l.ffn_up_exps->ne[1]*l.ffn_up_exps->nb[1];
ggml_build_forward_expand(gf, cpy1);
ggml_build_forward_expand(gf, cpy2);
} else {
ggml_build_forward_expand(gf, cpy1);
}
}
}
}
int n_threads = std::thread::hardware_concurrency()/2;
if (n_threads == 0) n_threads = 1;
llama_graph_compute(lctx, gf, n_threads);
llama_synchronize(&lctx);
ggml_backend_sched_reset(lctx.sched);
ggml_graph_clear(gf);
ggml_free(ctx);
*/
}
struct llama_context * llama_new_context_with_model(